[
  {
    "instance_id": "id1413",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3690,
        "sum_squared": 13616100,
        "min": 3690,
        "max": 3690,
        "mean": 3690.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 18,
        "sum_squared": 324,
        "min": 18,
        "max": 18,
        "mean": 18.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 18,
        "sum_squared": 324,
        "min": 18,
        "max": 18,
        "mean": 18.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.469085216522217,
        "sum_squared": 19.97272267253743,
        "min": 4.469085216522217,
        "max": 4.469085216522217,
        "mean": 4.469085216522217,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.36363636363636365,
        "sum_squared": 0.1322314049586777,
        "min": 0.36363636363636365,
        "max": 0.36363636363636365,
        "mean": 0.36363636363636365,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.35294117647058826,
        "sum_squared": 0.12456747404844293,
        "min": 0.35294117647058826,
        "max": 0.35294117647058826,
        "mean": 0.35294117647058826,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.11764705882352941,
        "sum_squared": 0.01384083044982699,
        "min": 0.11764705882352941,
        "max": 0.11764705882352941,
        "mean": 0.11764705882352941,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 18,
        "sum_squared": 324,
        "min": 18,
        "max": 18,
        "mean": 18.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 75,
        "sum_squared": 5625,
        "min": 75,
        "max": 75,
        "mean": 75.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1332",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3547,
        "sum_squared": 12581209,
        "min": 3547,
        "max": 3547,
        "mean": 3547.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 61,
        "sum_squared": 3721,
        "min": 61,
        "max": 61,
        "mean": 61.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 61,
        "sum_squared": 3721,
        "min": 61,
        "max": 61,
        "mean": 61.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.5164103507995605,
        "sum_squared": 30.43078315840853,
        "min": 5.5164103507995605,
        "max": 5.5164103507995605,
        "mean": 5.5164103507995605,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.12903225806451613,
        "sum_squared": 0.016649323621227886,
        "min": 0.12903225806451613,
        "max": 0.12903225806451613,
        "mean": 0.12903225806451613,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.09302325581395349,
        "sum_squared": 0.00865332612222823,
        "min": 0.09302325581395349,
        "max": 0.09302325581395349,
        "mean": 0.09302325581395349,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0588235294117647,
        "sum_squared": 0.0034602076124567466,
        "min": 0.0588235294117647,
        "max": 0.0588235294117647,
        "mean": 0.0588235294117647,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 61,
        "sum_squared": 3721,
        "min": 61,
        "max": 61,
        "mean": 61.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 242,
        "sum_squared": 58564,
        "min": 242,
        "max": 242,
        "mean": 242.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1123",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3585,
        "sum_squared": 12852225,
        "min": 3585,
        "max": 3585,
        "mean": 3585.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 56,
        "sum_squared": 3136,
        "min": 56,
        "max": 56,
        "mean": 56.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 56,
        "sum_squared": 3136,
        "min": 56,
        "max": 56,
        "mean": 56.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 7.400083303451538,
        "sum_squared": 54.76123289802223,
        "min": 7.400083303451538,
        "max": 7.400083303451538,
        "mean": 7.400083303451538,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.125,
        "sum_squared": 0.015625,
        "min": 0.125,
        "max": 0.125,
        "mean": 0.125,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.12,
        "sum_squared": 0.0144,
        "min": 0.12,
        "max": 0.12,
        "mean": 0.12,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.0588235294117647,
        "sum_squared": 0.0034602076124567466,
        "min": 0.0588235294117647,
        "max": 0.0588235294117647,
        "mean": 0.0588235294117647,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 56,
        "sum_squared": 3136,
        "min": 56,
        "max": 56,
        "mean": 56.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 255,
        "sum_squared": 65025,
        "min": 255,
        "max": 255,
        "mean": 255.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1514",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3672,
        "sum_squared": 13483584,
        "min": 3672,
        "max": 3672,
        "mean": 3672.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 89,
        "sum_squared": 7921,
        "min": 89,
        "max": 89,
        "mean": 89.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 89,
        "sum_squared": 7921,
        "min": 89,
        "max": 89,
        "mean": 89.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 7.491742372512817,
        "sum_squared": 56.12620377610398,
        "min": 7.491742372512817,
        "max": 7.491742372512817,
        "mean": 7.491742372512817,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07272727272727272,
        "sum_squared": 0.005289256198347107,
        "min": 0.07272727272727272,
        "max": 0.07272727272727272,
        "mean": 0.07272727272727272,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.05,
        "sum_squared": 0.0025000000000000005,
        "min": 0.05,
        "max": 0.05,
        "mean": 0.05,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.02352941176470588,
        "sum_squared": 0.0005536332179930794,
        "min": 0.02352941176470588,
        "max": 0.02352941176470588,
        "mean": 0.02352941176470588,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 89,
        "sum_squared": 7921,
        "min": 89,
        "max": 89,
        "mean": 89.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 413,
        "sum_squared": 170569,
        "min": 413,
        "max": 413,
        "mean": 413.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1340",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3208,
        "sum_squared": 10291264,
        "min": 3208,
        "max": 3208,
        "mean": 3208.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.8559958934783936,
        "sum_squared": 23.58069611747902,
        "min": 4.8559958934783936,
        "max": 4.8559958934783936,
        "mean": 4.8559958934783936,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.23529411764705882,
        "sum_squared": 0.05536332179930796,
        "min": 0.23529411764705882,
        "max": 0.23529411764705882,
        "mean": 0.23529411764705882,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2,
        "sum_squared": 0.04000000000000001,
        "min": 0.2,
        "max": 0.2,
        "mean": 0.2,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.22224546620451535,
        "sum_squared": 0.04939304724846238,
        "min": 0.22224546620451535,
        "max": 0.22224546620451535,
        "mean": 0.22224546620451535,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 1.6483752567447546e-308,
        "sum_squared": 0.0,
        "min": 1.6483752567447546e-308,
        "max": 1.6483752567447546e-308,
        "mean": 1.6483752567447546e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 41,
        "sum_squared": 1681,
        "min": 41,
        "max": 41,
        "mean": 41.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1281",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3776,
        "sum_squared": 14258176,
        "min": 3776,
        "max": 3776,
        "mean": 3776.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 7.271522283554077,
        "sum_squared": 52.8750363202235,
        "min": 7.271522283554077,
        "max": 7.271522283554077,
        "mean": 7.271522283554077,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.06666666666666667,
        "sum_squared": 0.0044444444444444444,
        "min": 0.06666666666666667,
        "max": 0.06666666666666667,
        "mean": 0.06666666666666667,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.04444444444444444,
        "sum_squared": 0.0019753086419753083,
        "min": 0.04444444444444444,
        "max": 0.04444444444444444,
        "mean": 0.04444444444444444,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.02083333333333333,
        "sum_squared": 0.0004340277777777776,
        "min": 0.02083333333333333,
        "max": 0.02083333333333333,
        "mean": 0.02083333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 519,
        "sum_squared": 269361,
        "min": 519,
        "max": 519,
        "mean": 519.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1234",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3500,
        "sum_squared": 12250000,
        "min": 3500,
        "max": 3500,
        "mean": 3500.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 57,
        "sum_squared": 3249,
        "min": 57,
        "max": 57,
        "mean": 57.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 57,
        "sum_squared": 3249,
        "min": 57,
        "max": 57,
        "mean": 57.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.436638355255127,
        "sum_squared": 29.557036605831172,
        "min": 5.436638355255127,
        "max": 5.436638355255127,
        "mean": 5.436638355255127,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.23809523809523814,
        "sum_squared": 0.056689342403628135,
        "min": 0.23809523809523814,
        "max": 0.23809523809523814,
        "mean": 0.23809523809523814,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1818181818181818,
        "sum_squared": 0.033057851239669415,
        "min": 0.1818181818181818,
        "max": 0.1818181818181818,
        "mean": 0.1818181818181818,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07999999999999999,
        "sum_squared": 0.006399999999999998,
        "min": 0.07999999999999999,
        "max": 0.07999999999999999,
        "mean": 0.07999999999999999,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 57,
        "sum_squared": 3249,
        "min": 57,
        "max": 57,
        "mean": 57.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 262,
        "sum_squared": 68644,
        "min": 262,
        "max": 262,
        "mean": 262.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1237",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3856,
        "sum_squared": 14868736,
        "min": 3856,
        "max": 3856,
        "mean": 3856.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.810382127761841,
        "sum_squared": 23.139776215090535,
        "min": 4.810382127761841,
        "max": 4.810382127761841,
        "mean": 4.810382127761841,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2857142857142857,
        "sum_squared": 0.08163265306122448,
        "min": 0.2857142857142857,
        "max": 0.2857142857142857,
        "mean": 0.2857142857142857,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.11111111111111109,
        "sum_squared": 0.012345679012345675,
        "min": 0.11111111111111109,
        "max": 0.11111111111111109,
        "mean": 0.11111111111111109,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 44,
        "sum_squared": 1936,
        "min": 44,
        "max": 44,
        "mean": 44.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1173",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3602,
        "sum_squared": 12974404,
        "min": 3602,
        "max": 3602,
        "mean": 3602.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 74,
        "sum_squared": 5476,
        "min": 74,
        "max": 74,
        "mean": 74.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 74,
        "sum_squared": 5476,
        "min": 74,
        "max": 74,
        "mean": 74.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 6.3067076206207275,
        "sum_squared": 39.77456101199556,
        "min": 6.3067076206207275,
        "max": 6.3067076206207275,
        "mean": 6.3067076206207275,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.42424242424242425,
        "sum_squared": 0.1799816345270891,
        "min": 0.42424242424242425,
        "max": 0.42424242424242425,
        "mean": 0.42424242424242425,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.27692307692307694,
        "sum_squared": 0.07668639053254439,
        "min": 0.27692307692307694,
        "max": 0.27692307692307694,
        "mean": 0.27692307692307694,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.16129032258064513,
        "sum_squared": 0.026014568158168563,
        "min": 0.16129032258064513,
        "max": 0.16129032258064513,
        "mean": 0.16129032258064513,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.016949152542372878,
        "sum_squared": 0.00028727377190462496,
        "min": 0.016949152542372878,
        "max": 0.016949152542372878,
        "mean": 0.016949152542372878,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 74,
        "sum_squared": 5476,
        "min": 74,
        "max": 74,
        "mean": 74.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 306,
        "sum_squared": 93636,
        "min": 306,
        "max": 306,
        "mean": 306.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1483",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3401,
        "sum_squared": 11566801,
        "min": 3401,
        "max": 3401,
        "mean": 3401.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 87,
        "sum_squared": 7569,
        "min": 87,
        "max": 87,
        "mean": 87.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 87,
        "sum_squared": 7569,
        "min": 87,
        "max": 87,
        "mean": 87.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 7.164652109146118,
        "sum_squared": 51.33223984509192,
        "min": 7.164652109146118,
        "max": 7.164652109146118,
        "mean": 7.164652109146118,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1090909090909091,
        "sum_squared": 0.011900826446280993,
        "min": 0.1090909090909091,
        "max": 0.1090909090909091,
        "mean": 0.1090909090909091,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0963855421686747,
        "sum_squared": 0.00929017273914937,
        "min": 0.0963855421686747,
        "max": 0.0963855421686747,
        "mean": 0.0963855421686747,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.06097560975609755,
        "sum_squared": 0.003718024985127899,
        "min": 0.06097560975609755,
        "max": 0.06097560975609755,
        "mean": 0.06097560975609755,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 87,
        "sum_squared": 7569,
        "min": 87,
        "max": 87,
        "mean": 87.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 461,
        "sum_squared": 212521,
        "min": 461,
        "max": 461,
        "mean": 461.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1474",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3799,
        "sum_squared": 14432401,
        "min": 3799,
        "max": 3799,
        "mean": 3799.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 3.4855751991271973,
        "sum_squared": 12.1492344687706,
        "min": 3.4855751991271973,
        "max": 3.4855751991271973,
        "mean": 3.4855751991271973,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.18181818181818182,
        "sum_squared": 0.03305785123966942,
        "min": 0.18181818181818182,
        "max": 0.18181818181818182,
        "mean": 0.18181818181818182,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.16666666666666666,
        "sum_squared": 0.027777777777777776,
        "min": 0.16666666666666666,
        "max": 0.16666666666666666,
        "mean": 0.16666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.10000000000000002,
        "sum_squared": 0.010000000000000004,
        "min": 0.10000000000000002,
        "max": 0.10000000000000002,
        "mean": 0.10000000000000002,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 48,
        "sum_squared": 2304,
        "min": 48,
        "max": 48,
        "mean": 48.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1571",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3856,
        "sum_squared": 14868736,
        "min": 3856,
        "max": 3856,
        "mean": 3856.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 44,
        "sum_squared": 1936,
        "min": 44,
        "max": 44,
        "mean": 44.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 44,
        "sum_squared": 1936,
        "min": 44,
        "max": 44,
        "mean": 44.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.46843695640564,
        "sum_squared": 29.903802746182976,
        "min": 5.46843695640564,
        "max": 5.46843695640564,
        "mean": 5.46843695640564,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.21428571428571427,
        "sum_squared": 0.04591836734693877,
        "min": 0.21428571428571427,
        "max": 0.21428571428571427,
        "mean": 0.21428571428571427,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.15,
        "sum_squared": 0.0225,
        "min": 0.15,
        "max": 0.15,
        "mean": 0.15,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.10256410256410255,
        "sum_squared": 0.010519395134779747,
        "min": 0.10256410256410255,
        "max": 0.10256410256410255,
        "mean": 0.10256410256410255,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 44,
        "sum_squared": 1936,
        "min": 44,
        "max": 44,
        "mean": 44.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 195,
        "sum_squared": 38025,
        "min": 195,
        "max": 195,
        "mean": 195.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1366",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3937,
        "sum_squared": 15499969,
        "min": 3937,
        "max": 3937,
        "mean": 3937.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 3.741358995437622,
        "sum_squared": 13.997767132742013,
        "min": 3.741358995437622,
        "max": 3.741358995437622,
        "mean": 3.741358995437622,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5714285714285714,
        "sum_squared": 0.32653061224489793,
        "min": 0.5714285714285714,
        "max": 0.5714285714285714,
        "mean": 0.5714285714285714,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5217391304347825,
        "sum_squared": 0.272211720226843,
        "min": 0.5217391304347825,
        "max": 0.5217391304347825,
        "mean": 0.5217391304347825,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3684210526315789,
        "sum_squared": 0.13573407202216065,
        "min": 0.3684210526315789,
        "max": 0.3684210526315789,
        "mean": 0.3684210526315789,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1442",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3743,
        "sum_squared": 14010049,
        "min": 3743,
        "max": 3743,
        "mean": 3743.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 7.3049139976501465,
        "sum_squared": 53.361768513065044,
        "min": 7.3049139976501465,
        "max": 7.3049139976501465,
        "mean": 7.3049139976501465,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.10169491525423728,
        "sum_squared": 0.010341855788566503,
        "min": 0.10169491525423728,
        "max": 0.10169491525423728,
        "mean": 0.10169491525423728,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.08791208791208792,
        "sum_squared": 0.007728535201062675,
        "min": 0.08791208791208792,
        "max": 0.08791208791208792,
        "mean": 0.08791208791208792,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.03225806451612903,
        "sum_squared": 0.0010405827263267429,
        "min": 0.03225806451612903,
        "max": 0.03225806451612903,
        "mean": 0.03225806451612903,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 494,
        "sum_squared": 244036,
        "min": 494,
        "max": 494,
        "mean": 494.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1529",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3490,
        "sum_squared": 12180100,
        "min": 3490,
        "max": 3490,
        "mean": 3490.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.315099239349365,
        "sum_squared": 28.2502799241322,
        "min": 5.315099239349365,
        "max": 5.315099239349365,
        "mean": 5.315099239349365,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.03225806451612903,
        "sum_squared": 0.0010405827263267429,
        "min": 0.03225806451612903,
        "max": 0.03225806451612903,
        "mean": 0.03225806451612903,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.024999999999999998,
        "sum_squared": 0.0006249999999999999,
        "min": 0.024999999999999998,
        "max": 0.024999999999999998,
        "mean": 0.024999999999999998,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.010752688172043006,
        "sum_squared": 0.00011562030292519358,
        "min": 0.010752688172043006,
        "max": 0.010752688172043006,
        "mean": 0.010752688172043006,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 456,
        "sum_squared": 207936,
        "min": 456,
        "max": 456,
        "mean": 456.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1209",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3695,
        "sum_squared": 13653025,
        "min": 3695,
        "max": 3695,
        "mean": 3695.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 35,
        "sum_squared": 1225,
        "min": 35,
        "max": 35,
        "mean": 35.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 35,
        "sum_squared": 1225,
        "min": 35,
        "max": 35,
        "mean": 35.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 6.618756532669067,
        "sum_squared": 43.807938038749455,
        "min": 6.618756532669067,
        "max": 6.618756532669067,
        "mean": 6.618756532669067,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.14814814814814814,
        "sum_squared": 0.02194787379972565,
        "min": 0.14814814814814814,
        "max": 0.14814814814814814,
        "mean": 0.14814814814814814,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.17647058823529413,
        "sum_squared": 0.031141868512110732,
        "min": 0.17647058823529413,
        "max": 0.17647058823529413,
        "mean": 0.17647058823529413,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.060606060606060615,
        "sum_squared": 0.0036730945821854925,
        "min": 0.060606060606060615,
        "max": 0.060606060606060615,
        "mean": 0.060606060606060615,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 35,
        "sum_squared": 1225,
        "min": 35,
        "max": 35,
        "mean": 35.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 174,
        "sum_squared": 30276,
        "min": 174,
        "max": 174,
        "mean": 174.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1408",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3451,
        "sum_squared": 11909401,
        "min": 3451,
        "max": 3451,
        "mean": 3451.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 3.7681643962860107,
        "sum_squared": 14.199062917437516,
        "min": 3.7681643962860107,
        "max": 3.7681643962860107,
        "mean": 3.7681643962860107,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.13333333333333333,
        "sum_squared": 0.017777777777777778,
        "min": 0.13333333333333333,
        "max": 0.13333333333333333,
        "mean": 0.13333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.125,
        "sum_squared": 0.015625,
        "min": 0.125,
        "max": 0.125,
        "mean": 0.125,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0909090909090909,
        "sum_squared": 0.008264462809917354,
        "min": 0.0909090909090909,
        "max": 0.0909090909090909,
        "mean": 0.0909090909090909,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1315",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3300,
        "sum_squared": 10890000,
        "min": 3300,
        "max": 3300,
        "mean": 3300.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 6.267944812774658,
        "sum_squared": 39.287132175988745,
        "min": 6.267944812774658,
        "max": 6.267944812774658,
        "mean": 6.267944812774658,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.16,
        "sum_squared": 0.0256,
        "min": 0.16,
        "max": 0.16,
        "mean": 0.16,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.12631578947368421,
        "sum_squared": 0.015955678670360112,
        "min": 0.12631578947368421,
        "max": 0.12631578947368421,
        "mean": 0.12631578947368421,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.04301075268817205,
        "sum_squared": 0.0018499248468030996,
        "min": 0.04301075268817205,
        "max": 0.04301075268817205,
        "mean": 0.04301075268817205,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.011111111111111112,
        "sum_squared": 0.0001234567901234568,
        "min": 0.011111111111111112,
        "max": 0.011111111111111112,
        "mean": 0.011111111111111112,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 505,
        "sum_squared": 255025,
        "min": 505,
        "max": 505,
        "mean": 505.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1421",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3962,
        "sum_squared": 15697444,
        "min": 3962,
        "max": 3962,
        "mean": 3962.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 36,
        "sum_squared": 1296,
        "min": 36,
        "max": 36,
        "mean": 36.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 36,
        "sum_squared": 1296,
        "min": 36,
        "max": 36,
        "mean": 36.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.54633641242981,
        "sum_squared": 20.66917477498515,
        "min": 4.54633641242981,
        "max": 4.54633641242981,
        "mean": 4.54633641242981,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.32258064516129037,
        "sum_squared": 0.10405827263267432,
        "min": 0.32258064516129037,
        "max": 0.32258064516129037,
        "mean": 0.32258064516129037,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.30769230769230765,
        "sum_squared": 0.09467455621301772,
        "min": 0.30769230769230765,
        "max": 0.30769230769230765,
        "mean": 0.30769230769230765,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.03448275862068965,
        "sum_squared": 0.0011890606420927462,
        "min": 0.03448275862068965,
        "max": 0.03448275862068965,
        "mean": 0.03448275862068965,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 36,
        "sum_squared": 1296,
        "min": 36,
        "max": 36,
        "mean": 36.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 154,
        "sum_squared": 23716,
        "min": 154,
        "max": 154,
        "mean": 154.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1308",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3823,
        "sum_squared": 14615329,
        "min": 3823,
        "max": 3823,
        "mean": 3823.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.533444881439209,
        "sum_squared": 30.61901225592578,
        "min": 5.533444881439209,
        "max": 5.533444881439209,
        "mean": 5.533444881439209,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.15384615384615385,
        "sum_squared": 0.02366863905325444,
        "min": 0.15384615384615385,
        "max": 0.15384615384615385,
        "mean": 0.15384615384615385,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.11764705882352941,
        "sum_squared": 0.01384083044982699,
        "min": 0.11764705882352941,
        "max": 0.11764705882352941,
        "mean": 0.11764705882352941,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.05263157894736841,
        "sum_squared": 0.002770083102493074,
        "min": 0.05263157894736841,
        "max": 0.05263157894736841,
        "mean": 0.05263157894736841,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 109,
        "sum_squared": 11881,
        "min": 109,
        "max": 109,
        "mean": 109.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1456",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3634,
        "sum_squared": 13205956,
        "min": 3634,
        "max": 3634,
        "mean": 3634.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 26,
        "sum_squared": 676,
        "min": 26,
        "max": 26,
        "mean": 26.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 26,
        "sum_squared": 676,
        "min": 26,
        "max": 26,
        "mean": 26.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.940703630447388,
        "sum_squared": 24.410552363915997,
        "min": 4.940703630447388,
        "max": 4.940703630447388,
        "mean": 4.940703630447388,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.31578947368421056,
        "sum_squared": 0.09972299168975071,
        "min": 0.31578947368421056,
        "max": 0.31578947368421056,
        "mean": 0.31578947368421056,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.12000000000000001,
        "sum_squared": 0.014400000000000003,
        "min": 0.12000000000000001,
        "max": 0.12000000000000001,
        "mean": 0.12000000000000001,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 26,
        "sum_squared": 676,
        "min": 26,
        "max": 26,
        "mean": 26.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 141,
        "sum_squared": 19881,
        "min": 141,
        "max": 141,
        "mean": 141.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1335",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3838,
        "sum_squared": 14730244,
        "min": 3838,
        "max": 3838,
        "mean": 3838.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 22,
        "sum_squared": 484,
        "min": 22,
        "max": 22,
        "mean": 22.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 22,
        "sum_squared": 484,
        "min": 22,
        "max": 22,
        "mean": 22.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.257296323776245,
        "sum_squared": 27.63916463599122,
        "min": 5.257296323776245,
        "max": 5.257296323776245,
        "mean": 5.257296323776245,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.21052631578947367,
        "sum_squared": 0.04432132963988919,
        "min": 0.21052631578947367,
        "max": 0.21052631578947367,
        "mean": 0.21052631578947367,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.05263157894736841,
        "sum_squared": 0.002770083102493074,
        "min": 0.05263157894736841,
        "max": 0.05263157894736841,
        "mean": 0.05263157894736841,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 22,
        "sum_squared": 484,
        "min": 22,
        "max": 22,
        "mean": 22.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 93,
        "sum_squared": 8649,
        "min": 93,
        "max": 93,
        "mean": 93.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1337",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3570,
        "sum_squared": 12744900,
        "min": 3570,
        "max": 3570,
        "mean": 3570.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 40,
        "sum_squared": 1600,
        "min": 40,
        "max": 40,
        "mean": 40.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 40,
        "sum_squared": 1600,
        "min": 40,
        "max": 40,
        "mean": 40.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.676121950149536,
        "sum_squared": 21.8661164926703,
        "min": 4.676121950149536,
        "max": 4.676121950149536,
        "mean": 4.676121950149536,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0,
        "sum_squared": 0.0,
        "min": 0.0,
        "max": 0.0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 40,
        "sum_squared": 1600,
        "min": 40,
        "max": 40,
        "mean": 40.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 227,
        "sum_squared": 51529,
        "min": 227,
        "max": 227,
        "mean": 227.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1114",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3332,
        "sum_squared": 11102224,
        "min": 3332,
        "max": 3332,
        "mean": 3332.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 21,
        "sum_squared": 441,
        "min": 21,
        "max": 21,
        "mean": 21.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 21,
        "sum_squared": 441,
        "min": 21,
        "max": 21,
        "mean": 21.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 4.507139444351196,
        "sum_squared": 20.31430597082641,
        "min": 4.507139444351196,
        "max": 4.507139444351196,
        "mean": 4.507139444351196,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.14285714285714285,
        "sum_squared": 0.02040816326530612,
        "min": 0.14285714285714285,
        "max": 0.14285714285714285,
        "mean": 0.14285714285714285,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.11764705882352941,
        "sum_squared": 0.01384083044982699,
        "min": 0.11764705882352941,
        "max": 0.11764705882352941,
        "mean": 0.11764705882352941,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.05000000000000001,
        "sum_squared": 0.002500000000000001,
        "min": 0.05000000000000001,
        "max": 0.05000000000000001,
        "mean": 0.05000000000000001,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 21,
        "sum_squared": 441,
        "min": 21,
        "max": 21,
        "mean": 21.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 97,
        "sum_squared": 9409,
        "min": 97,
        "max": 97,
        "mean": 97.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1198",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3749,
        "sum_squared": 14055001,
        "min": 3749,
        "max": 3749,
        "mean": 3749.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 48,
        "sum_squared": 2304,
        "min": 48,
        "max": 48,
        "mean": 48.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 48,
        "sum_squared": 2304,
        "min": 48,
        "max": 48,
        "mean": 48.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 6.074464321136475,
        "sum_squared": 36.89911678876001,
        "min": 6.074464321136475,
        "max": 6.074464321136475,
        "mean": 6.074464321136475,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.47368421052631576,
        "sum_squared": 0.22437673130193903,
        "min": 0.47368421052631576,
        "max": 0.47368421052631576,
        "mean": 0.47368421052631576,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.2916666666666667,
        "sum_squared": 0.08506944444444446,
        "min": 0.2916666666666667,
        "max": 0.2916666666666667,
        "mean": 0.2916666666666667,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.24390243902439027,
        "sum_squared": 0.05948839976204641,
        "min": 0.24390243902439027,
        "max": 0.24390243902439027,
        "mean": 0.24390243902439027,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.05263157894736841,
        "sum_squared": 0.002770083102493074,
        "min": 0.05263157894736841,
        "max": 0.05263157894736841,
        "mean": 0.05263157894736841,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 48,
        "sum_squared": 2304,
        "min": 48,
        "max": 48,
        "mean": 48.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 219,
        "sum_squared": 47961,
        "min": 219,
        "max": 219,
        "mean": 219.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1178",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3283,
        "sum_squared": 10778089,
        "min": 3283,
        "max": 3283,
        "mean": 3283.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 6.8471519947052,
        "sum_squared": 46.8834904385954,
        "min": 6.8471519947052,
        "max": 6.8471519947052,
        "mean": 6.8471519947052,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.09302325581395349,
        "sum_squared": 0.00865332612222823,
        "min": 0.09302325581395349,
        "max": 0.09302325581395349,
        "mean": 0.09302325581395349,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.07228915662650602,
        "sum_squared": 0.005225722165771519,
        "min": 0.07228915662650602,
        "max": 0.07228915662650602,
        "mean": 0.07228915662650602,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.03125,
        "sum_squared": 0.0009765625,
        "min": 0.03125,
        "max": 0.03125,
        "mean": 0.03125,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 465,
        "sum_squared": 216225,
        "min": 465,
        "max": 465,
        "mean": 465.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1117",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3680,
        "sum_squared": 13542400,
        "min": 3680,
        "max": 3680,
        "mean": 3680.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 6.341818809509277,
        "sum_squared": 40.21866581264567,
        "min": 6.341818809509277,
        "max": 6.341818809509277,
        "mean": 6.341818809509277,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.03508771929824561,
        "sum_squared": 0.0012311480455524776,
        "min": 0.03508771929824561,
        "max": 0.03508771929824561,
        "mean": 0.03508771929824561,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.04545454545454545,
        "sum_squared": 0.0020661157024793385,
        "min": 0.04545454545454545,
        "max": 0.04545454545454545,
        "mean": 0.04545454545454545,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.010869565217391302,
        "sum_squared": 0.00011814744801512282,
        "min": 0.010869565217391302,
        "max": 0.010869565217391302,
        "mean": 0.010869565217391302,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 515,
        "sum_squared": 265225,
        "min": 515,
        "max": 515,
        "mean": 515.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1273",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3980,
        "sum_squared": 15840400,
        "min": 3980,
        "max": 3980,
        "mean": 3980.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 39,
        "sum_squared": 1521,
        "min": 39,
        "max": 39,
        "mean": 39.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 39,
        "sum_squared": 1521,
        "min": 39,
        "max": 39,
        "mean": 39.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.486333131790161,
        "sum_squared": 30.099851232978438,
        "min": 5.486333131790161,
        "max": 5.486333131790161,
        "mean": 5.486333131790161,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5499999999999999,
        "sum_squared": 0.30249999999999994,
        "min": 0.5499999999999999,
        "max": 0.5499999999999999,
        "mean": 0.5499999999999999,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5531914893617021,
        "sum_squared": 0.30602082390221824,
        "min": 0.5531914893617021,
        "max": 0.5531914893617021,
        "mean": 0.5531914893617021,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.32432432432432434,
        "sum_squared": 0.10518626734842952,
        "min": 0.32432432432432434,
        "max": 0.32432432432432434,
        "mean": 0.32432432432432434,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.14705882352941177,
        "sum_squared": 0.021626297577854673,
        "min": 0.14705882352941177,
        "max": 0.14705882352941177,
        "mean": 0.14705882352941177,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 39,
        "sum_squared": 1521,
        "min": 39,
        "max": 39,
        "mean": 39.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 199,
        "sum_squared": 39601,
        "min": 199,
        "max": 199,
        "mean": 199.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1180",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 4016,
        "sum_squared": 16128256,
        "min": 4016,
        "max": 4016,
        "mean": 4016.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 7.587849140167236,
        "sum_squared": 57.57545457393667,
        "min": 7.587849140167236,
        "max": 7.587849140167236,
        "mean": 7.587849140167236,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.07407407407407407,
        "sum_squared": 0.0054869684499314125,
        "min": 0.07407407407407407,
        "max": 0.07407407407407407,
        "mean": 0.07407407407407407,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.04597701149425288,
        "sum_squared": 0.0021138855859426614,
        "min": 0.04597701149425288,
        "max": 0.04597701149425288,
        "mean": 0.04597701149425288,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.021276595744680854,
        "sum_squared": 0.0004526935264825714,
        "min": 0.021276595744680854,
        "max": 0.021276595744680854,
        "mean": 0.021276595744680854,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 489,
        "sum_squared": 239121,
        "min": 489,
        "max": 489,
        "mean": 489.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1261",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3942,
        "sum_squared": 15539364,
        "min": 3942,
        "max": 3942,
        "mean": 3942.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 20,
        "sum_squared": 400,
        "min": 20,
        "max": 20,
        "mean": 20.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 20,
        "sum_squared": 400,
        "min": 20,
        "max": 20,
        "mean": 20.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.83321213722229,
        "sum_squared": 23.359939563392857,
        "min": 4.83321213722229,
        "max": 4.83321213722229,
        "mean": 4.83321213722229,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0,
        "sum_squared": 0.0,
        "min": 0.0,
        "max": 0.0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 20,
        "sum_squared": 400,
        "min": 20,
        "max": 20,
        "mean": 20.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 79,
        "sum_squared": 6241,
        "min": 79,
        "max": 79,
        "mean": 79.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1290",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3793,
        "sum_squared": 14386849,
        "min": 3793,
        "max": 3793,
        "mean": 3793.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 37,
        "sum_squared": 1369,
        "min": 37,
        "max": 37,
        "mean": 37.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 37,
        "sum_squared": 1369,
        "min": 37,
        "max": 37,
        "mean": 37.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.079019069671631,
        "sum_squared": 25.79643471008808,
        "min": 5.079019069671631,
        "max": 5.079019069671631,
        "mean": 5.079019069671631,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3846153846153846,
        "sum_squared": 0.14792899408284022,
        "min": 0.3846153846153846,
        "max": 0.3846153846153846,
        "mean": 0.3846153846153846,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.34285714285714286,
        "sum_squared": 0.11755102040816327,
        "min": 0.34285714285714286,
        "max": 0.34285714285714286,
        "mean": 0.34285714285714286,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.15151515151515155,
        "sum_squared": 0.02295684113865933,
        "min": 0.15151515151515155,
        "max": 0.15151515151515155,
        "mean": 0.15151515151515155,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 37,
        "sum_squared": 1369,
        "min": 37,
        "max": 37,
        "mean": 37.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 167,
        "sum_squared": 27889,
        "min": 167,
        "max": 167,
        "mean": 167.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1202",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3802,
        "sum_squared": 14455204,
        "min": 3802,
        "max": 3802,
        "mean": 3802.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 66,
        "sum_squared": 4356,
        "min": 66,
        "max": 66,
        "mean": 66.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 66,
        "sum_squared": 4356,
        "min": 66,
        "max": 66,
        "mean": 66.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 5.524125814437866,
        "sum_squared": 30.51596601373882,
        "min": 5.524125814437866,
        "max": 5.524125814437866,
        "mean": 5.524125814437866,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.34285714285714286,
        "sum_squared": 0.11755102040816327,
        "min": 0.34285714285714286,
        "max": 0.34285714285714286,
        "mean": 0.34285714285714286,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.25925925925925924,
        "sum_squared": 0.0672153635116598,
        "min": 0.25925925925925924,
        "max": 0.25925925925925924,
        "mean": 0.25925925925925924,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.14545454545454545,
        "sum_squared": 0.021157024793388428,
        "min": 0.14545454545454545,
        "max": 0.14545454545454545,
        "mean": 0.14545454545454545,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.03846153846153846,
        "sum_squared": 0.001479289940828402,
        "min": 0.03846153846153846,
        "max": 0.03846153846153846,
        "mean": 0.03846153846153846,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 66,
        "sum_squared": 4356,
        "min": 66,
        "max": 66,
        "mean": 66.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 265,
        "sum_squared": 70225,
        "min": 265,
        "max": 265,
        "mean": 265.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1463",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3841,
        "sum_squared": 14753281,
        "min": 3841,
        "max": 3841,
        "mean": 3841.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 33,
        "sum_squared": 1089,
        "min": 33,
        "max": 33,
        "mean": 33.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 33,
        "sum_squared": 1089,
        "min": 33,
        "max": 33,
        "mean": 33.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.199464797973633,
        "sum_squared": 27.03443418536699,
        "min": 5.199464797973633,
        "max": 5.199464797973633,
        "mean": 5.199464797973633,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3333333333333333,
        "sum_squared": 0.1111111111111111,
        "min": 0.3333333333333333,
        "max": 0.3333333333333333,
        "mean": 0.3333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.19999999999999998,
        "sum_squared": 0.039999999999999994,
        "min": 0.19999999999999998,
        "max": 0.19999999999999998,
        "mean": 0.19999999999999998,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.12903225806451615,
        "sum_squared": 0.016649323621227893,
        "min": 0.12903225806451615,
        "max": 0.12903225806451615,
        "mean": 0.12903225806451615,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.03571428571428572,
        "sum_squared": 0.001275510204081633,
        "min": 0.03571428571428572,
        "max": 0.03571428571428572,
        "mean": 0.03571428571428572,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 33,
        "sum_squared": 1089,
        "min": 33,
        "max": 33,
        "mean": 33.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 138,
        "sum_squared": 19044,
        "min": 138,
        "max": 138,
        "mean": 138.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1112",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3755,
        "sum_squared": 14100025,
        "min": 3755,
        "max": 3755,
        "mean": 3755.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 4.911074876785278,
        "sum_squared": 24.118656445391537,
        "min": 4.911074876785278,
        "max": 4.911074876785278,
        "mean": 4.911074876785278,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.36363636363636365,
        "sum_squared": 0.1322314049586777,
        "min": 0.36363636363636365,
        "max": 0.36363636363636365,
        "mean": 0.36363636363636365,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.42857142857142855,
        "sum_squared": 0.18367346938775508,
        "min": 0.42857142857142855,
        "max": 0.42857142857142855,
        "mean": 0.42857142857142855,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.21739130434782608,
        "sum_squared": 0.04725897920604915,
        "min": 0.21739130434782608,
        "max": 0.21739130434782608,
        "mean": 0.21739130434782608,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.10000000000000002,
        "sum_squared": 0.010000000000000004,
        "min": 0.10000000000000002,
        "max": 0.10000000000000002,
        "mean": 0.10000000000000002,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 111,
        "sum_squared": 12321,
        "min": 111,
        "max": 111,
        "mean": 111.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1293",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3759,
        "sum_squared": 14130081,
        "min": 3759,
        "max": 3759,
        "mean": 3759.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 31,
        "sum_squared": 961,
        "min": 31,
        "max": 31,
        "mean": 31.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 31,
        "sum_squared": 961,
        "min": 31,
        "max": 31,
        "mean": 31.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 6.1989710330963135,
        "sum_squared": 38.427241869167176,
        "min": 6.1989710330963135,
        "max": 6.1989710330963135,
        "mean": 6.1989710330963135,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1111111111111111,
        "sum_squared": 0.012345679012345678,
        "min": 0.1111111111111111,
        "max": 0.1111111111111111,
        "mean": 0.1111111111111111,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0909090909090909,
        "sum_squared": 0.008264462809917354,
        "min": 0.0909090909090909,
        "max": 0.0909090909090909,
        "mean": 0.0909090909090909,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.08333333333333333,
        "sum_squared": 0.006944444444444444,
        "min": 0.08333333333333333,
        "max": 0.08333333333333333,
        "mean": 0.08333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 31,
        "sum_squared": 961,
        "min": 31,
        "max": 31,
        "mean": 31.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 115,
        "sum_squared": 13225,
        "min": 115,
        "max": 115,
        "mean": 115.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1256",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3371,
        "sum_squared": 11363641,
        "min": 3371,
        "max": 3371,
        "mean": 3371.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 6.593636512756348,
        "sum_squared": 43.47604246235369,
        "min": 6.593636512756348,
        "max": 6.593636512756348,
        "mean": 6.593636512756348,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07547169811320754,
        "sum_squared": 0.005695977216091135,
        "min": 0.07547169811320754,
        "max": 0.07547169811320754,
        "mean": 0.07547169811320754,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07228915662650602,
        "sum_squared": 0.005225722165771519,
        "min": 0.07228915662650602,
        "max": 0.07228915662650602,
        "mean": 0.07228915662650602,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0449438202247191,
        "sum_squared": 0.0020199469763918696,
        "min": 0.0449438202247191,
        "max": 0.0449438202247191,
        "mean": 0.0449438202247191,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 488,
        "sum_squared": 238144,
        "min": 488,
        "max": 488,
        "mean": 488.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1499",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3815,
        "sum_squared": 14554225,
        "min": 3815,
        "max": 3815,
        "mean": 3815.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 31,
        "sum_squared": 961,
        "min": 31,
        "max": 31,
        "mean": 31.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 31,
        "sum_squared": 961,
        "min": 31,
        "max": 31,
        "mean": 31.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 6.197995662689209,
        "sum_squared": 38.41515023471425,
        "min": 6.197995662689209,
        "max": 6.197995662689209,
        "mean": 6.197995662689209,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.37037037037037035,
        "sum_squared": 0.1371742112482853,
        "min": 0.37037037037037035,
        "max": 0.37037037037037035,
        "mean": 0.37037037037037035,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.33333333333333337,
        "sum_squared": 0.11111111111111113,
        "min": 0.33333333333333337,
        "max": 0.33333333333333337,
        "mean": 0.33333333333333337,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.15384615384615385,
        "sum_squared": 0.02366863905325444,
        "min": 0.15384615384615385,
        "max": 0.15384615384615385,
        "mean": 0.15384615384615385,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.043478260869565216,
        "sum_squared": 0.0018903591682419658,
        "min": 0.043478260869565216,
        "max": 0.043478260869565216,
        "mean": 0.043478260869565216,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 31,
        "sum_squared": 961,
        "min": 31,
        "max": 31,
        "mean": 31.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 140,
        "sum_squared": 19600,
        "min": 140,
        "max": 140,
        "mean": 140.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1419",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3290,
        "sum_squared": 10824100,
        "min": 3290,
        "max": 3290,
        "mean": 3290.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 63,
        "sum_squared": 3969,
        "min": 63,
        "max": 63,
        "mean": 63.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 63,
        "sum_squared": 3969,
        "min": 63,
        "max": 63,
        "mean": 63.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.552546501159668,
        "sum_squared": 30.83077264754047,
        "min": 5.552546501159668,
        "max": 5.552546501159668,
        "mean": 5.552546501159668,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1111111111111111,
        "sum_squared": 0.012345679012345678,
        "min": 0.1111111111111111,
        "max": 0.1111111111111111,
        "mean": 0.1111111111111111,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.11538461538461538,
        "sum_squared": 0.01331360946745562,
        "min": 0.11538461538461538,
        "max": 0.11538461538461538,
        "mean": 0.11538461538461538,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.05263157894736841,
        "sum_squared": 0.002770083102493074,
        "min": 0.05263157894736841,
        "max": 0.05263157894736841,
        "mean": 0.05263157894736841,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 63,
        "sum_squared": 3969,
        "min": 63,
        "max": 63,
        "mean": 63.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 292,
        "sum_squared": 85264,
        "min": 292,
        "max": 292,
        "mean": 292.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1103",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3664,
        "sum_squared": 13424896,
        "min": 3664,
        "max": 3664,
        "mean": 3664.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 37,
        "sum_squared": 1369,
        "min": 37,
        "max": 37,
        "mean": 37.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 37,
        "sum_squared": 1369,
        "min": 37,
        "max": 37,
        "mean": 37.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 5.292325258255005,
        "sum_squared": 28.008706639163904,
        "min": 5.292325258255005,
        "max": 5.292325258255005,
        "mean": 5.292325258255005,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.09090909090909091,
        "sum_squared": 0.008264462809917356,
        "min": 0.09090909090909091,
        "max": 0.09090909090909091,
        "mean": 0.09090909090909091,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.06666666666666667,
        "sum_squared": 0.0044444444444444444,
        "min": 0.06666666666666667,
        "max": 0.06666666666666667,
        "mean": 0.06666666666666667,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.030303030303030304,
        "sum_squared": 0.0009182736455463729,
        "min": 0.030303030303030304,
        "max": 0.030303030303030304,
        "mean": 0.030303030303030304,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 37,
        "sum_squared": 1369,
        "min": 37,
        "max": 37,
        "mean": 37.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 179,
        "sum_squared": 32041,
        "min": 179,
        "max": 179,
        "mean": 179.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1331",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3448,
        "sum_squared": 11888704,
        "min": 3448,
        "max": 3448,
        "mean": 3448.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 55,
        "sum_squared": 3025,
        "min": 55,
        "max": 55,
        "mean": 55.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 55,
        "sum_squared": 3025,
        "min": 55,
        "max": 55,
        "mean": 55.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 3.8882830142974854,
        "sum_squared": 15.118744799274339,
        "min": 3.8882830142974854,
        "max": 3.8882830142974854,
        "mean": 3.8882830142974854,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.15384615384615385,
        "sum_squared": 0.02366863905325444,
        "min": 0.15384615384615385,
        "max": 0.15384615384615385,
        "mean": 0.15384615384615385,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.11320754716981132,
        "sum_squared": 0.012815948736205056,
        "min": 0.11320754716981132,
        "max": 0.11320754716981132,
        "mean": 0.11320754716981132,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07272727272727272,
        "sum_squared": 0.005289256198347107,
        "min": 0.07272727272727272,
        "max": 0.07272727272727272,
        "mean": 0.07272727272727272,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 55,
        "sum_squared": 3025,
        "min": 55,
        "max": 55,
        "mean": 55.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 266,
        "sum_squared": 70756,
        "min": 266,
        "max": 266,
        "mean": 266.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1410",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3370,
        "sum_squared": 11356900,
        "min": 3370,
        "max": 3370,
        "mean": 3370.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 6.9792046546936035,
        "sum_squared": 48.70929761209686,
        "min": 6.9792046546936035,
        "max": 6.9792046546936035,
        "mean": 6.9792046546936035,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0425531914893617,
        "sum_squared": 0.0018107741059302852,
        "min": 0.0425531914893617,
        "max": 0.0425531914893617,
        "mean": 0.0425531914893617,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.04395604395604395,
        "sum_squared": 0.0019321338002656681,
        "min": 0.04395604395604395,
        "max": 0.04395604395604395,
        "mean": 0.04395604395604395,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.02105263157894737,
        "sum_squared": 0.0004432132963988921,
        "min": 0.02105263157894737,
        "max": 0.02105263157894737,
        "mean": 0.02105263157894737,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 500,
        "sum_squared": 250000,
        "min": 500,
        "max": 500,
        "mean": 500.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1516",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3666,
        "sum_squared": 13439556,
        "min": 3666,
        "max": 3666,
        "mean": 3666.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.867039680480957,
        "sum_squared": 34.42215461233809,
        "min": 5.867039680480957,
        "max": 5.867039680480957,
        "mean": 5.867039680480957,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.125,
        "sum_squared": 0.015625,
        "min": 0.125,
        "max": 0.125,
        "mean": 0.125,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0909090909090909,
        "sum_squared": 0.008264462809917354,
        "min": 0.0909090909090909,
        "max": 0.0909090909090909,
        "mean": 0.0909090909090909,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.041666666666666664,
        "sum_squared": 0.001736111111111111,
        "min": 0.041666666666666664,
        "max": 0.041666666666666664,
        "mean": 0.041666666666666664,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 126,
        "sum_squared": 15876,
        "min": 126,
        "max": 126,
        "mean": 126.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1449",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3330,
        "sum_squared": 11088900,
        "min": 3330,
        "max": 3330,
        "mean": 3330.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 56,
        "sum_squared": 3136,
        "min": 56,
        "max": 56,
        "mean": 56.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 56,
        "sum_squared": 3136,
        "min": 56,
        "max": 56,
        "mean": 56.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.920541286468506,
        "sum_squared": 24.21172655184114,
        "min": 4.920541286468506,
        "max": 4.920541286468506,
        "mean": 4.920541286468506,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.16666666666666666,
        "sum_squared": 0.027777777777777776,
        "min": 0.16666666666666666,
        "max": 0.16666666666666666,
        "mean": 0.16666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1568627450980392,
        "sum_squared": 0.024605920799692427,
        "min": 0.1568627450980392,
        "max": 0.1568627450980392,
        "mean": 0.1568627450980392,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07547169811320754,
        "sum_squared": 0.005695977216091135,
        "min": 0.07547169811320754,
        "max": 0.07547169811320754,
        "mean": 0.07547169811320754,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 56,
        "sum_squared": 3136,
        "min": 56,
        "max": 56,
        "mean": 56.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 263,
        "sum_squared": 69169,
        "min": 263,
        "max": 263,
        "mean": 263.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1411",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3852,
        "sum_squared": 14837904,
        "min": 3852,
        "max": 3852,
        "mean": 3852.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 73,
        "sum_squared": 5329,
        "min": 73,
        "max": 73,
        "mean": 73.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 73,
        "sum_squared": 5329,
        "min": 73,
        "max": 73,
        "mean": 73.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 6.237070560455322,
        "sum_squared": 38.90104917609847,
        "min": 6.237070560455322,
        "max": 6.237070560455322,
        "mean": 6.237070560455322,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.11428571428571428,
        "sum_squared": 0.013061224489795917,
        "min": 0.11428571428571428,
        "max": 0.11428571428571428,
        "mean": 0.11428571428571428,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07843137254901959,
        "sum_squared": 0.006151480199923104,
        "min": 0.07843137254901959,
        "max": 0.07843137254901959,
        "mean": 0.07843137254901959,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.03448275862068965,
        "sum_squared": 0.0011890606420927462,
        "min": 0.03448275862068965,
        "max": 0.03448275862068965,
        "mean": 0.03448275862068965,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 73,
        "sum_squared": 5329,
        "min": 73,
        "max": 73,
        "mean": 73.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 276,
        "sum_squared": 76176,
        "min": 276,
        "max": 276,
        "mean": 276.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1495",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3745,
        "sum_squared": 14025025,
        "min": 3745,
        "max": 3745,
        "mean": 3745.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 9.59612488746643,
        "sum_squared": 92.08561285585262,
        "min": 9.59612488746643,
        "max": 9.59612488746643,
        "mean": 9.59612488746643,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1754385964912281,
        "sum_squared": 0.03077870113881195,
        "min": 0.1754385964912281,
        "max": 0.1754385964912281,
        "mean": 0.1754385964912281,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1518987341772152,
        "sum_squared": 0.023073225444640282,
        "min": 0.1518987341772152,
        "max": 0.1518987341772152,
        "mean": 0.1518987341772152,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.08888888888888888,
        "sum_squared": 0.007901234567901233,
        "min": 0.08888888888888888,
        "max": 0.08888888888888888,
        "mean": 0.08888888888888888,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.03448275862068965,
        "sum_squared": 0.0011890606420927462,
        "min": 0.03448275862068965,
        "max": 0.03448275862068965,
        "mean": 0.03448275862068965,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 430,
        "sum_squared": 184900,
        "min": 430,
        "max": 430,
        "mean": 430.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1445",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3219,
        "sum_squared": 10361961,
        "min": 3219,
        "max": 3219,
        "mean": 3219.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.315507888793945,
        "sum_squared": 18.623608338242775,
        "min": 4.315507888793945,
        "max": 4.315507888793945,
        "mean": 4.315507888793945,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2857142857142857,
        "sum_squared": 0.08163265306122448,
        "min": 0.2857142857142857,
        "max": 0.2857142857142857,
        "mean": 0.2857142857142857,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2222222222222222,
        "sum_squared": 0.04938271604938271,
        "min": 0.2222222222222222,
        "max": 0.2222222222222222,
        "mean": 0.2222222222222222,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.16666666666666669,
        "sum_squared": 0.027777777777777783,
        "min": 0.16666666666666669,
        "max": 0.16666666666666669,
        "mean": 0.16666666666666669,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 89,
        "sum_squared": 7921,
        "min": 89,
        "max": 89,
        "mean": 89.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1507",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4295,
        "sum_squared": 18447025,
        "min": 4295,
        "max": 4295,
        "mean": 4295.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 21,
        "sum_squared": 441,
        "min": 21,
        "max": 21,
        "mean": 21.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 21,
        "sum_squared": 441,
        "min": 21,
        "max": 21,
        "mean": 21.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 5.032076597213745,
        "sum_squared": 25.321794880226264,
        "min": 5.032076597213745,
        "max": 5.032076597213745,
        "mean": 5.032076597213745,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3076923076923077,
        "sum_squared": 0.09467455621301776,
        "min": 0.3076923076923077,
        "max": 0.3076923076923077,
        "mean": 0.3076923076923077,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.23529411764705882,
        "sum_squared": 0.05536332179930796,
        "min": 0.23529411764705882,
        "max": 0.23529411764705882,
        "mean": 0.23529411764705882,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.11111111111111109,
        "sum_squared": 0.012345679012345675,
        "min": 0.11111111111111109,
        "max": 0.11111111111111109,
        "mean": 0.11111111111111109,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 21,
        "sum_squared": 441,
        "min": 21,
        "max": 21,
        "mean": 21.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 79,
        "sum_squared": 6241,
        "min": 79,
        "max": 79,
        "mean": 79.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1414",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4160,
        "sum_squared": 17305600,
        "min": 4160,
        "max": 4160,
        "mean": 4160.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 46,
        "sum_squared": 2116,
        "min": 46,
        "max": 46,
        "mean": 46.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 46,
        "sum_squared": 2116,
        "min": 46,
        "max": 46,
        "mean": 46.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 6.306368350982666,
        "sum_squared": 39.77028177827583,
        "min": 6.306368350982666,
        "max": 6.306368350982666,
        "mean": 6.306368350982666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.15384615384615385,
        "sum_squared": 0.02366863905325444,
        "min": 0.15384615384615385,
        "max": 0.15384615384615385,
        "mean": 0.15384615384615385,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.13333333333333333,
        "sum_squared": 0.017777777777777778,
        "min": 0.13333333333333333,
        "max": 0.13333333333333333,
        "mean": 0.13333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.054054054054054064,
        "sum_squared": 0.0029218407596785984,
        "min": 0.054054054054054064,
        "max": 0.054054054054054064,
        "mean": 0.054054054054054064,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 46,
        "sum_squared": 2116,
        "min": 46,
        "max": 46,
        "mean": 46.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 180,
        "sum_squared": 32400,
        "min": 180,
        "max": 180,
        "mean": 180.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1246",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4047,
        "sum_squared": 16378209,
        "min": 4047,
        "max": 4047,
        "mean": 4047.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 23,
        "sum_squared": 529,
        "min": 23,
        "max": 23,
        "mean": 23.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 23,
        "sum_squared": 529,
        "min": 23,
        "max": 23,
        "mean": 23.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 4.636913061141968,
        "sum_squared": 21.500962736588974,
        "min": 4.636913061141968,
        "max": 4.636913061141968,
        "mean": 4.636913061141968,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1111111111111111,
        "sum_squared": 0.012345679012345678,
        "min": 0.1111111111111111,
        "max": 0.1111111111111111,
        "mean": 0.1111111111111111,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.047619047619047616,
        "sum_squared": 0.0022675736961451243,
        "min": 0.047619047619047616,
        "max": 0.047619047619047616,
        "mean": 0.047619047619047616,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 23,
        "sum_squared": 529,
        "min": 23,
        "max": 23,
        "mean": 23.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 105,
        "sum_squared": 11025,
        "min": 105,
        "max": 105,
        "mean": 105.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1555",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3297,
        "sum_squared": 10870209,
        "min": 3297,
        "max": 3297,
        "mean": 3297.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 7.155101299285889,
        "sum_squared": 51.19547460304261,
        "min": 7.155101299285889,
        "max": 7.155101299285889,
        "mean": 7.155101299285889,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.12987012987012986,
        "sum_squared": 0.016866250632484394,
        "min": 0.12987012987012986,
        "max": 0.12987012987012986,
        "mean": 0.12987012987012986,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07547169811320754,
        "sum_squared": 0.005695977216091135,
        "min": 0.07547169811320754,
        "max": 0.07547169811320754,
        "mean": 0.07547169811320754,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.06315789473684211,
        "sum_squared": 0.003988919667590028,
        "min": 0.06315789473684211,
        "max": 0.06315789473684211,
        "mean": 0.06315789473684211,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 100,
        "sum_squared": 10000,
        "min": 100,
        "max": 100,
        "mean": 100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 494,
        "sum_squared": 244036,
        "min": 494,
        "max": 494,
        "mean": 494.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  }
]