[
  {
    "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": 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": 2.224378824234009,
        "sum_squared": 4.947861153700671,
        "min": 2.224378824234009,
        "max": 2.224378824234009,
        "mean": 2.224378824234009,
        "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": 0.05714285714285714,
        "sum_squared": 0.0032653061224489793,
        "min": 0.05714285714285714,
        "max": 0.05714285714285714,
        "mean": 0.05714285714285714,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.02941176470588235,
        "sum_squared": 0.0008650519031141867,
        "min": 0.02941176470588235,
        "max": 0.02941176470588235,
        "mean": 0.02941176470588235,
        "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": 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": 181,
        "sum_squared": 32761,
        "min": 181,
        "max": 181,
        "mean": 181.0,
        "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": 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": 2.45698618888855,
        "sum_squared": 6.0367811323890805,
        "min": 2.45698618888855,
        "max": 2.45698618888855,
        "mean": 2.45698618888855,
        "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": 0.028571428571428577,
        "sum_squared": 0.0008163265306122453,
        "min": 0.028571428571428577,
        "max": 0.028571428571428577,
        "mean": 0.028571428571428577,
        "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": 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": 206,
        "sum_squared": 42436,
        "min": 206,
        "max": 206,
        "mean": 206.0,
        "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": 73,
        "sum_squared": 5329,
        "min": 73,
        "max": 73,
        "mean": 73.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "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": "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": 2.6848134994506836,
        "sum_squared": 7.208223526832626,
        "min": 2.6848134994506836,
        "max": 2.6848134994506836,
        "mean": 2.6848134994506836,
        "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": 0.08571428571428572,
        "sum_squared": 0.007346938775510205,
        "min": 0.08571428571428572,
        "max": 0.08571428571428572,
        "mean": 0.08571428571428572,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "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": "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": 73,
        "sum_squared": 5329,
        "min": 73,
        "max": 73,
        "mean": 73.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 347,
        "sum_squared": 120409,
        "min": 347,
        "max": 347,
        "mean": 347.0,
        "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": 25,
        "sum_squared": 625,
        "min": 25,
        "max": 25,
        "mean": 25.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 25,
        "sum_squared": 625,
        "min": 25,
        "max": 25,
        "mean": 25.0,
        "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": 2.1628000736236572,
        "sum_squared": 4.677704158466497,
        "min": 2.1628000736236572,
        "max": 2.1628000736236572,
        "mean": 2.1628000736236572,
        "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": 25,
        "sum_squared": 625,
        "min": 25,
        "max": 25,
        "mean": 25.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 113,
        "sum_squared": 12769,
        "min": 113,
        "max": 113,
        "mean": 113.0,
        "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": 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": 2.534818410873413,
        "sum_squared": 6.425304376102815,
        "min": 2.534818410873413,
        "max": 2.534818410873413,
        "mean": 2.534818410873413,
        "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": 0.13953488372093023,
        "sum_squared": 0.019469983775013522,
        "min": 0.13953488372093023,
        "max": 0.13953488372093023,
        "mean": 0.13953488372093023,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.09836065573770492,
        "sum_squared": 0.009674818597151302,
        "min": 0.09836065573770492,
        "max": 0.09836065573770492,
        "mean": 0.09836065573770492,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.10909090909090911,
        "sum_squared": 0.011900826446280996,
        "min": 0.10909090909090911,
        "max": 0.10909090909090911,
        "mean": 0.10909090909090911,
        "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": 277,
        "sum_squared": 76729,
        "min": 277,
        "max": 277,
        "mean": 277.0,
        "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": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "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": 2.664872407913208,
        "sum_squared": 7.101544950457139,
        "min": 2.664872407913208,
        "max": 2.664872407913208,
        "mean": 2.664872407913208,
        "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": 0.11764705882352941,
        "sum_squared": 0.01384083044982699,
        "min": 0.11764705882352941,
        "max": 0.11764705882352941,
        "mean": 0.11764705882352941,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.09523809523809525,
        "sum_squared": 0.0090702947845805,
        "min": 0.09523809523809525,
        "max": 0.09523809523809525,
        "mean": 0.09523809523809525,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.048780487804878044,
        "sum_squared": 0.0023795359904818557,
        "min": 0.048780487804878044,
        "max": 0.048780487804878044,
        "mean": 0.048780487804878044,
        "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": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 221,
        "sum_squared": 48841,
        "min": 221,
        "max": 221,
        "mean": 221.0,
        "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": 38,
        "sum_squared": 1444,
        "min": 38,
        "max": 38,
        "mean": 38.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 38,
        "sum_squared": 1444,
        "min": 38,
        "max": 38,
        "mean": 38.0,
        "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": 2.3789918422698975,
        "sum_squared": 5.659602185586721,
        "min": 2.3789918422698975,
        "max": 2.3789918422698975,
        "mean": 2.3789918422698975,
        "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": 0.12121212121212122,
        "sum_squared": 0.014692378328741967,
        "min": 0.12121212121212122,
        "max": 0.12121212121212122,
        "mean": 0.12121212121212122,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.09523809523809523,
        "sum_squared": 0.009070294784580497,
        "min": 0.09523809523809523,
        "max": 0.09523809523809523,
        "mean": 0.09523809523809523,
        "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": 38,
        "sum_squared": 1444,
        "min": 38,
        "max": 38,
        "mean": 38.0,
        "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": "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": 53,
        "sum_squared": 2809,
        "min": 53,
        "max": 53,
        "mean": 53.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 53,
        "sum_squared": 2809,
        "min": 53,
        "max": 53,
        "mean": 53.0,
        "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": 2.4179036617279053,
        "sum_squared": 5.846258117397213,
        "min": 2.4179036617279053,
        "max": 2.4179036617279053,
        "mean": 2.4179036617279053,
        "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": 53,
        "sum_squared": 2809,
        "min": 53,
        "max": 53,
        "mean": 53.0,
        "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": "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": 29,
        "sum_squared": 841,
        "min": 29,
        "max": 29,
        "mean": 29.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 29,
        "sum_squared": 841,
        "min": 29,
        "max": 29,
        "mean": 29.0,
        "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": 2.2988924980163574,
        "sum_squared": 5.284906717435888,
        "min": 2.2988924980163574,
        "max": 2.2988924980163574,
        "mean": 2.2988924980163574,
        "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": 0.06896551724137931,
        "sum_squared": 0.0047562425683709865,
        "min": 0.06896551724137931,
        "max": 0.06896551724137931,
        "mean": 0.06896551724137931,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.05555555555555555,
        "sum_squared": 0.0030864197530864196,
        "min": 0.05555555555555555,
        "max": 0.05555555555555555,
        "mean": 0.05555555555555555,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "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": "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": 29,
        "sum_squared": 841,
        "min": 29,
        "max": 29,
        "mean": 29.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 159,
        "sum_squared": 25281,
        "min": 159,
        "max": 159,
        "mean": 159.0,
        "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": 34,
        "sum_squared": 1156,
        "min": 34,
        "max": 34,
        "mean": 34.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 34,
        "sum_squared": 1156,
        "min": 34,
        "max": 34,
        "mean": 34.0,
        "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": 2.68803334236145,
        "sum_squared": 7.225523249646869,
        "min": 2.68803334236145,
        "max": 2.68803334236145,
        "mean": 2.68803334236145,
        "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": 0.03225806451612903,
        "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": 34,
        "sum_squared": 1156,
        "min": 34,
        "max": 34,
        "mean": 34.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 174,
        "sum_squared": 30276,
        "min": 174,
        "max": 174,
        "mean": 174.0,
        "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": 78,
        "sum_squared": 6084,
        "min": 78,
        "max": 78,
        "mean": 78.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 78,
        "sum_squared": 6084,
        "min": 78,
        "max": 78,
        "mean": 78.0,
        "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": 3.1861729621887207,
        "sum_squared": 10.151698144982447,
        "min": 3.1861729621887207,
        "max": 3.1861729621887207,
        "mean": 3.1861729621887207,
        "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": 0.11538461538461536,
        "sum_squared": 0.013313609467455616,
        "min": 0.11538461538461536,
        "max": 0.11538461538461536,
        "mean": 0.11538461538461536,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1176470588235294,
        "sum_squared": 0.013840830449826987,
        "min": 0.1176470588235294,
        "max": 0.1176470588235294,
        "mean": 0.1176470588235294,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07352941176470587,
        "sum_squared": 0.0054065743944636665,
        "min": 0.07352941176470587,
        "max": 0.07352941176470587,
        "mean": 0.07352941176470587,
        "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": 78,
        "sum_squared": 6084,
        "min": 78,
        "max": 78,
        "mean": 78.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 353,
        "sum_squared": 124609,
        "min": 353,
        "max": 353,
        "mean": 353.0,
        "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": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "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": 2.320415496826172,
        "sum_squared": 5.38432807791105,
        "min": 2.320415496826172,
        "max": 2.320415496826172,
        "mean": 2.320415496826172,
        "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": 0.05714285714285714,
        "sum_squared": 0.0032653061224489793,
        "min": 0.05714285714285714,
        "max": 0.05714285714285714,
        "mean": 0.05714285714285714,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "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_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": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 247,
        "sum_squared": 61009,
        "min": 247,
        "max": 247,
        "mean": 247.0,
        "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": 25,
        "sum_squared": 625,
        "min": 25,
        "max": 25,
        "mean": 25.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 25,
        "sum_squared": 625,
        "min": 25,
        "max": 25,
        "mean": 25.0,
        "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": 2.424398183822632,
        "sum_squared": 5.877706553722476,
        "min": 2.424398183822632,
        "max": 2.424398183822632,
        "mean": 2.424398183822632,
        "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": 0.043478260869565216,
        "sum_squared": 0.0018903591682419658,
        "min": 0.043478260869565216,
        "max": 0.043478260869565216,
        "mean": 0.043478260869565216,
        "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": 25,
        "sum_squared": 625,
        "min": 25,
        "max": 25,
        "mean": 25.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 129,
        "sum_squared": 16641,
        "min": 129,
        "max": 129,
        "mean": 129.0,
        "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": 52,
        "sum_squared": 2704,
        "min": 52,
        "max": 52,
        "mean": 52.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 52,
        "sum_squared": 2704,
        "min": 52,
        "max": 52,
        "mean": 52.0,
        "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": 2.5365147590637207,
        "sum_squared": 6.433907122948085,
        "min": 2.5365147590637207,
        "max": 2.5365147590637207,
        "mean": 2.5365147590637207,
        "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": 0.04,
        "sum_squared": 0.0016,
        "min": 0.04,
        "max": 0.04,
        "mean": 0.04,
        "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": 52,
        "sum_squared": 2704,
        "min": 52,
        "max": 52,
        "mean": 52.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 246,
        "sum_squared": 60516,
        "min": 246,
        "max": 246,
        "mean": 246.0,
        "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": 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": 2.8354954719543457,
        "sum_squared": 8.040034571473598,
        "min": 2.8354954719543457,
        "max": 2.8354954719543457,
        "mean": 2.8354954719543457,
        "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": 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": 127,
        "sum_squared": 16129,
        "min": 127,
        "max": 127,
        "mean": 127.0,
        "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": 58,
        "sum_squared": 3364,
        "min": 58,
        "max": 58,
        "mean": 58.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 58,
        "sum_squared": 3364,
        "min": 58,
        "max": 58,
        "mean": 58.0,
        "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": 3.5131776332855225,
        "sum_squared": 12.342417083017665,
        "min": 3.5131776332855225,
        "max": 3.5131776332855225,
        "mean": 3.5131776332855225,
        "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": 0.05405405405405406,
        "sum_squared": 0.002921840759678598,
        "min": 0.05405405405405406,
        "max": 0.05405405405405406,
        "mean": 0.05405405405405406,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.0392156862745098,
        "sum_squared": 0.0015378700499807767,
        "min": 0.0392156862745098,
        "max": 0.0392156862745098,
        "mean": 0.0392156862745098,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.018867924528301883,
        "sum_squared": 0.00035599857600569583,
        "min": 0.018867924528301883,
        "max": 0.018867924528301883,
        "mean": 0.018867924528301883,
        "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": 58,
        "sum_squared": 3364,
        "min": 58,
        "max": 58,
        "mean": 58.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 291,
        "sum_squared": 84681,
        "min": 291,
        "max": 291,
        "mean": 291.0,
        "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": 49,
        "sum_squared": 2401,
        "min": 49,
        "max": 49,
        "mean": 49.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 49,
        "sum_squared": 2401,
        "min": 49,
        "max": 49,
        "mean": 49.0,
        "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": 2.062563896179199,
        "sum_squared": 4.2541698258219185,
        "min": 2.062563896179199,
        "max": 2.062563896179199,
        "mean": 2.062563896179199,
        "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": 0.05555555555555555,
        "sum_squared": 0.0030864197530864196,
        "min": 0.05555555555555555,
        "max": 0.05555555555555555,
        "mean": 0.05555555555555555,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "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": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.04444444444444445,
        "sum_squared": 0.0019753086419753096,
        "min": 0.04444444444444445,
        "max": 0.04444444444444445,
        "mean": 0.04444444444444445,
        "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": 49,
        "sum_squared": 2401,
        "min": 49,
        "max": 49,
        "mean": 49.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 247,
        "sum_squared": 61009,
        "min": 247,
        "max": 247,
        "mean": 247.0,
        "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": 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": 2.6501736640930176,
        "sum_squared": 7.02342044985221,
        "min": 2.6501736640930176,
        "max": 2.6501736640930176,
        "mean": 2.6501736640930176,
        "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": 0.1568627450980392,
        "sum_squared": 0.024605920799692427,
        "min": 0.1568627450980392,
        "max": 0.1568627450980392,
        "mean": 0.1568627450980392,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.14634146341463414,
        "sum_squared": 0.0214158239143367,
        "min": 0.14634146341463414,
        "max": 0.14634146341463414,
        "mean": 0.14634146341463414,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.05063291139240507,
        "sum_squared": 0.0025636917160711432,
        "min": 0.05063291139240507,
        "max": 0.05063291139240507,
        "mean": 0.05063291139240507,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.013157894736842106,
        "sum_squared": 0.0001731301939058172,
        "min": 0.013157894736842106,
        "max": 0.013157894736842106,
        "mean": 0.013157894736842106,
        "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": 419,
        "sum_squared": 175561,
        "min": 419,
        "max": 419,
        "mean": 419.0,
        "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": 43,
        "sum_squared": 1849,
        "min": 43,
        "max": 43,
        "mean": 43.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 43,
        "sum_squared": 1849,
        "min": 43,
        "max": 43,
        "mean": 43.0,
        "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": 2.644747734069824,
        "sum_squared": 6.99469057686747,
        "min": 2.644747734069824,
        "max": 2.644747734069824,
        "mean": 2.644747734069824,
        "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": 0.06060606060606061,
        "sum_squared": 0.0036730945821854917,
        "min": 0.06060606060606061,
        "max": 0.06060606060606061,
        "mean": 0.06060606060606061,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.11764705882352942,
        "sum_squared": 0.013840830449826992,
        "min": 0.11764705882352942,
        "max": 0.11764705882352942,
        "mean": 0.11764705882352942,
        "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": 43,
        "sum_squared": 1849,
        "min": 43,
        "max": 43,
        "mean": 43.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 214,
        "sum_squared": 45796,
        "min": 214,
        "max": 214,
        "mean": 214.0,
        "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": 49,
        "sum_squared": 2401,
        "min": 49,
        "max": 49,
        "mean": 49.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 49,
        "sum_squared": 2401,
        "min": 49,
        "max": 49,
        "mean": 49.0,
        "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": 2.734177350997925,
        "sum_squared": 7.475725786710029,
        "min": 2.734177350997925,
        "max": 2.734177350997925,
        "mean": 2.734177350997925,
        "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": 0.09090909090909091,
        "sum_squared": 0.008264462809917356,
        "min": 0.09090909090909091,
        "max": 0.09090909090909091,
        "mean": 0.09090909090909091,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0625,
        "sum_squared": 0.00390625,
        "min": 0.0625,
        "max": 0.0625,
        "mean": 0.0625,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.028571428571428577,
        "sum_squared": 0.0008163265306122453,
        "min": 0.028571428571428577,
        "max": 0.028571428571428577,
        "mean": 0.028571428571428577,
        "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": 49,
        "sum_squared": 2401,
        "min": 49,
        "max": 49,
        "mean": 49.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 204,
        "sum_squared": 41616,
        "min": 204,
        "max": 204,
        "mean": 204.0,
        "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": 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": 3.148038625717163,
        "sum_squared": 9.910147189007205,
        "min": 3.148038625717163,
        "max": 3.148038625717163,
        "mean": 3.148038625717163,
        "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": 0.15789473684210528,
        "sum_squared": 0.024930747922437678,
        "min": 0.15789473684210528,
        "max": 0.15789473684210528,
        "mean": 0.15789473684210528,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07894736842105263,
        "sum_squared": 0.006232686980609418,
        "min": 0.07894736842105263,
        "max": 0.07894736842105263,
        "mean": 0.07894736842105263,
        "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": 221,
        "sum_squared": 48841,
        "min": 221,
        "max": 221,
        "mean": 221.0,
        "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": 60,
        "sum_squared": 3600,
        "min": 60,
        "max": 60,
        "mean": 60.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 60,
        "sum_squared": 3600,
        "min": 60,
        "max": 60,
        "mean": 60.0,
        "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": 2.742969036102295,
        "sum_squared": 7.523879133015953,
        "min": 2.742969036102295,
        "max": 2.742969036102295,
        "mean": 2.742969036102295,
        "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": 0.12121212121212122,
        "sum_squared": 0.014692378328741967,
        "min": 0.12121212121212122,
        "max": 0.12121212121212122,
        "mean": 0.12121212121212122,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.08163265306122448,
        "sum_squared": 0.006663890045814243,
        "min": 0.08163265306122448,
        "max": 0.08163265306122448,
        "mean": 0.08163265306122448,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.03921568627450981,
        "sum_squared": 0.001537870049980777,
        "min": 0.03921568627450981,
        "max": 0.03921568627450981,
        "mean": 0.03921568627450981,
        "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": 60,
        "sum_squared": 3600,
        "min": 60,
        "max": 60,
        "mean": 60.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 273,
        "sum_squared": 74529,
        "min": 273,
        "max": 273,
        "mean": 273.0,
        "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": 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": 2.3568341732025146,
        "sum_squared": 5.554667319975181,
        "min": 2.3568341732025146,
        "max": 2.3568341732025146,
        "mean": 2.3568341732025146,
        "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": 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": 211,
        "sum_squared": 44521,
        "min": 211,
        "max": 211,
        "mean": 211.0,
        "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": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "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": 8.123433113098145,
        "sum_squared": 65.99016554297941,
        "min": 8.123433113098145,
        "max": 8.123433113098145,
        "mean": 8.123433113098145,
        "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": 0.06666666666666667,
        "sum_squared": 0.0044444444444444444,
        "min": 0.06666666666666667,
        "max": 0.06666666666666667,
        "mean": 0.06666666666666667,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "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": 0.021276595744680854,
        "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": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 235,
        "sum_squared": 55225,
        "min": 235,
        "max": 235,
        "mean": 235.0,
        "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": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "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": 2.5073401927948,
        "sum_squared": 6.286754842404264,
        "min": 2.5073401927948,
        "max": 2.5073401927948,
        "mean": 2.5073401927948,
        "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,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "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": "valid"
        },
        "count": 1,
        "sum": 0.07692307692307693,
        "sum_squared": 0.00591715976331361,
        "min": 0.07692307692307693,
        "max": 0.07692307692307693,
        "mean": 0.07692307692307693,
        "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": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 68,
        "sum_squared": 4624,
        "min": 68,
        "max": 68,
        "mean": 68.0,
        "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": 28,
        "sum_squared": 784,
        "min": 28,
        "max": 28,
        "mean": 28.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 28,
        "sum_squared": 784,
        "min": 28,
        "max": 28,
        "mean": 28.0,
        "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": 2.235477924346924,
        "sum_squared": 4.997361550242431,
        "min": 2.235477924346924,
        "max": 2.235477924346924,
        "mean": 2.235477924346924,
        "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": 0.2,
        "sum_squared": 0.04000000000000001,
        "min": 0.2,
        "max": 0.2,
        "mean": 0.2,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.23076923076923078,
        "sum_squared": 0.053254437869822494,
        "min": 0.23076923076923078,
        "max": 0.23076923076923078,
        "mean": 0.23076923076923078,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "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": "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": 28,
        "sum_squared": 784,
        "min": 28,
        "max": 28,
        "mean": 28.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 127,
        "sum_squared": 16129,
        "min": 127,
        "max": 127,
        "mean": 127.0,
        "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": 36,
        "sum_squared": 1296,
        "min": 36,
        "max": 36,
        "mean": 36.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "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": "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": 2.8593881130218506,
        "sum_squared": 8.17610038089066,
        "min": 2.8593881130218506,
        "max": 2.8593881130218506,
        "mean": 2.8593881130218506,
        "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,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "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": "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": 36,
        "sum_squared": 1296,
        "min": 36,
        "max": 36,
        "mean": 36.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 175,
        "sum_squared": 30625,
        "min": 175,
        "max": 175,
        "mean": 175.0,
        "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": 60,
        "sum_squared": 3600,
        "min": 60,
        "max": 60,
        "mean": 60.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 60,
        "sum_squared": 3600,
        "min": 60,
        "max": 60,
        "mean": 60.0,
        "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": 2.7947824001312256,
        "sum_squared": 7.810808664083254,
        "min": 2.7947824001312256,
        "max": 2.7947824001312256,
        "mean": 2.7947824001312256,
        "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": 0.11764705882352941,
        "sum_squared": 0.01384083044982699,
        "min": 0.11764705882352941,
        "max": 0.11764705882352941,
        "mean": 0.11764705882352941,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "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_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.06896551724137931,
        "sum_squared": 0.0047562425683709865,
        "min": 0.06896551724137931,
        "max": 0.06896551724137931,
        "mean": 0.06896551724137931,
        "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": 60,
        "sum_squared": 3600,
        "min": 60,
        "max": 60,
        "mean": 60.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 319,
        "sum_squared": 101761,
        "min": 319,
        "max": 319,
        "mean": 319.0,
        "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": 3.432880401611328,
        "sum_squared": 11.784667851767153,
        "min": 3.432880401611328,
        "max": 3.432880401611328,
        "mean": 3.432880401611328,
        "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": 0.08695652173913043,
        "sum_squared": 0.007561436672967863,
        "min": 0.08695652173913043,
        "max": 0.08695652173913043,
        "mean": 0.08695652173913043,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.05063291139240507,
        "sum_squared": 0.0025636917160711432,
        "min": 0.05063291139240507,
        "max": 0.05063291139240507,
        "mean": 0.05063291139240507,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.02272727272727273,
        "sum_squared": 0.0005165289256198349,
        "min": 0.02272727272727273,
        "max": 0.02272727272727273,
        "mean": 0.02272727272727273,
        "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": 439,
        "sum_squared": 192721,
        "min": 439,
        "max": 439,
        "mean": 439.0,
        "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": 82,
        "sum_squared": 6724,
        "min": 82,
        "max": 82,
        "mean": 82.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 82,
        "sum_squared": 6724,
        "min": 82,
        "max": 82,
        "mean": 82.0,
        "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": 3.303844928741455,
        "sum_squared": 10.91539131317063,
        "min": 3.303844928741455,
        "max": 3.303844928741455,
        "mean": 3.303844928741455,
        "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": 0.028169014084507046,
        "sum_squared": 0.0007934933544931563,
        "min": 0.028169014084507046,
        "max": 0.028169014084507046,
        "mean": 0.028169014084507046,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.05333333333333334,
        "sum_squared": 0.002844444444444445,
        "min": 0.05333333333333334,
        "max": 0.05333333333333334,
        "mean": 0.05333333333333334,
        "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": 82,
        "sum_squared": 6724,
        "min": 82,
        "max": 82,
        "mean": 82.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 390,
        "sum_squared": 152100,
        "min": 390,
        "max": 390,
        "mean": 390.0,
        "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": 84,
        "sum_squared": 7056,
        "min": 84,
        "max": 84,
        "mean": 84.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 84,
        "sum_squared": 7056,
        "min": 84,
        "max": 84,
        "mean": 84.0,
        "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": 3.0943946838378906,
        "sum_squared": 9.575278459364199,
        "min": 3.0943946838378906,
        "max": 3.0943946838378906,
        "mean": 3.0943946838378906,
        "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": 0.07407407407407407,
        "sum_squared": 0.0054869684499314125,
        "min": 0.07407407407407407,
        "max": 0.07407407407407407,
        "mean": 0.07407407407407407,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0810810810810811,
        "sum_squared": 0.006574141709276848,
        "min": 0.0810810810810811,
        "max": 0.0810810810810811,
        "mean": 0.0810810810810811,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.039473684210526314,
        "sum_squared": 0.0015581717451523544,
        "min": 0.039473684210526314,
        "max": 0.039473684210526314,
        "mean": 0.039473684210526314,
        "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": 84,
        "sum_squared": 7056,
        "min": 84,
        "max": 84,
        "mean": 84.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 399,
        "sum_squared": 159201,
        "min": 399,
        "max": 399,
        "mean": 399.0,
        "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": 22,
        "sum_squared": 484,
        "min": 22,
        "max": 22,
        "mean": 22.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "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": "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": 2.391998052597046,
        "sum_squared": 5.72165468362806,
        "min": 2.391998052597046,
        "max": 2.391998052597046,
        "mean": 2.391998052597046,
        "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": 0.2857142857142857,
        "sum_squared": 0.08163265306122448,
        "min": 0.2857142857142857,
        "max": 0.2857142857142857,
        "mean": 0.2857142857142857,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.32,
        "sum_squared": 0.1024,
        "min": 0.32,
        "max": 0.32,
        "mean": 0.32,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.2777777777777778,
        "sum_squared": 0.0771604938271605,
        "min": 0.2777777777777778,
        "max": 0.2777777777777778,
        "mean": 0.2777777777777778,
        "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": 22,
        "sum_squared": 484,
        "min": 22,
        "max": 22,
        "mean": 22.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 86,
        "sum_squared": 7396,
        "min": 86,
        "max": 86,
        "mean": 86.0,
        "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": 67,
        "sum_squared": 4489,
        "min": 67,
        "max": 67,
        "mean": 67.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 67,
        "sum_squared": 4489,
        "min": 67,
        "max": 67,
        "mean": 67.0,
        "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": 2.9864821434020996,
        "sum_squared": 8.919075592859599,
        "min": 2.9864821434020996,
        "max": 2.9864821434020996,
        "mean": 2.9864821434020996,
        "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": 0.041666666666666664,
        "sum_squared": 0.001736111111111111,
        "min": 0.041666666666666664,
        "max": 0.041666666666666664,
        "mean": 0.041666666666666664,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.031746031746031744,
        "sum_squared": 0.0010078105316200553,
        "min": 0.031746031746031744,
        "max": 0.031746031746031744,
        "mean": 0.031746031746031744,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.03174603174603174,
        "sum_squared": 0.0010078105316200549,
        "min": 0.03174603174603174,
        "max": 0.03174603174603174,
        "mean": 0.03174603174603174,
        "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": 67,
        "sum_squared": 4489,
        "min": 67,
        "max": 67,
        "mean": 67.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 336,
        "sum_squared": 112896,
        "min": 336,
        "max": 336,
        "mean": 336.0,
        "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": 33,
        "sum_squared": 1089,
        "min": 33,
        "max": 33,
        "mean": 33.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "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": "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": 2.12265944480896,
        "sum_squared": 4.505683118636682,
        "min": 2.12265944480896,
        "max": 2.12265944480896,
        "mean": 2.12265944480896,
        "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": 0.06451612903225806,
        "sum_squared": 0.004162330905306971,
        "min": 0.06451612903225806,
        "max": 0.06451612903225806,
        "mean": 0.06451612903225806,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.10526315789473685,
        "sum_squared": 0.011080332409972301,
        "min": 0.10526315789473685,
        "max": 0.10526315789473685,
        "mean": 0.10526315789473685,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "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": "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": 33,
        "sum_squared": 1089,
        "min": 33,
        "max": 33,
        "mean": 33.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 167,
        "sum_squared": 27889,
        "min": 167,
        "max": 167,
        "mean": 167.0,
        "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": 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": 2.480572462081909,
        "sum_squared": 6.153239739639105,
        "min": 2.480572462081909,
        "max": 2.480572462081909,
        "mean": 2.480572462081909,
        "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": 0.05263157894736841,
        "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": 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": 102,
        "sum_squared": 10404,
        "min": 102,
        "max": 102,
        "mean": 102.0,
        "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": 50,
        "sum_squared": 2500,
        "min": 50,
        "max": 50,
        "mean": 50.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 50,
        "sum_squared": 2500,
        "min": 50,
        "max": 50,
        "mean": 50.0,
        "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": 2.6337101459503174,
        "sum_squared": 6.936429132881642,
        "min": 2.6337101459503174,
        "max": 2.6337101459503174,
        "mean": 2.6337101459503174,
        "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": 0.14634146341463414,
        "sum_squared": 0.0214158239143367,
        "min": 0.14634146341463414,
        "max": 0.14634146341463414,
        "mean": 0.14634146341463414,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.09523809523809525,
        "sum_squared": 0.0090702947845805,
        "min": 0.09523809523809525,
        "max": 0.09523809523809525,
        "mean": 0.09523809523809525,
        "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": 50,
        "sum_squared": 2500,
        "min": 50,
        "max": 50,
        "mean": 50.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 225,
        "sum_squared": 50625,
        "min": 225,
        "max": 225,
        "mean": 225.0,
        "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": 24,
        "sum_squared": 576,
        "min": 24,
        "max": 24,
        "mean": 24.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 24,
        "sum_squared": 576,
        "min": 24,
        "max": 24,
        "mean": 24.0,
        "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": 2.279789924621582,
        "sum_squared": 5.197442100406079,
        "min": 2.279789924621582,
        "max": 2.279789924621582,
        "mean": 2.279789924621582,
        "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": 0.041666666666666664,
        "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": 24,
        "sum_squared": 576,
        "min": 24,
        "max": 24,
        "mean": 24.0,
        "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": "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": 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": 2.745532751083374,
        "sum_squared": 7.53795008727144,
        "min": 2.745532751083374,
        "max": 2.745532751083374,
        "mean": 2.745532751083374,
        "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": 0.05263157894736842,
        "sum_squared": 0.0027700831024930744,
        "min": 0.05263157894736842,
        "max": 0.05263157894736842,
        "mean": 0.05263157894736842,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.026315789473684213,
        "sum_squared": 0.0006925207756232688,
        "min": 0.026315789473684213,
        "max": 0.026315789473684213,
        "mean": 0.026315789473684213,
        "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": 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": 209,
        "sum_squared": 43681,
        "min": 209,
        "max": 209,
        "mean": 209.0,
        "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": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "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": 2.799980401992798,
        "sum_squared": 7.83989025154375,
        "min": 2.799980401992798,
        "max": 2.799980401992798,
        "mean": 2.799980401992798,
        "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,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "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": "valid"
        },
        "count": 1,
        "sum": 0.022222222222222227,
        "sum_squared": 0.0004938271604938274,
        "min": 0.022222222222222227,
        "max": 0.022222222222222227,
        "mean": 0.022222222222222227,
        "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": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 249,
        "sum_squared": 62001,
        "min": 249,
        "max": 249,
        "mean": 249.0,
        "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": 54,
        "sum_squared": 2916,
        "min": 54,
        "max": 54,
        "mean": 54.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 54,
        "sum_squared": 2916,
        "min": 54,
        "max": 54,
        "mean": 54.0,
        "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": 2.6191627979278564,
        "sum_squared": 6.860013762049277,
        "min": 2.6191627979278564,
        "max": 2.6191627979278564,
        "mean": 2.6191627979278564,
        "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": 0.05555555555555555,
        "sum_squared": 0.0030864197530864196,
        "min": 0.05555555555555555,
        "max": 0.05555555555555555,
        "mean": 0.05555555555555555,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0392156862745098,
        "sum_squared": 0.0015378700499807767,
        "min": 0.0392156862745098,
        "max": 0.0392156862745098,
        "mean": 0.0392156862745098,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.03773584905660377,
        "sum_squared": 0.0014239943040227838,
        "min": 0.03773584905660377,
        "max": 0.03773584905660377,
        "mean": 0.03773584905660377,
        "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": 54,
        "sum_squared": 2916,
        "min": 54,
        "max": 54,
        "mean": 54.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 249,
        "sum_squared": 62001,
        "min": 249,
        "max": 249,
        "mean": 249.0,
        "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": 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": 1.4848790168762207,
        "sum_squared": 2.2048656947592917,
        "min": 1.4848790168762207,
        "max": 1.4848790168762207,
        "mean": 1.4848790168762207,
        "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": 0.041666666666666664,
        "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": 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": 121,
        "sum_squared": 14641,
        "min": 121,
        "max": 121,
        "mean": 121.0,
        "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": 59,
        "sum_squared": 3481,
        "min": 59,
        "max": 59,
        "mean": 59.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 59,
        "sum_squared": 3481,
        "min": 59,
        "max": 59,
        "mean": 59.0,
        "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": 3.130516767501831,
        "sum_squared": 9.800135231610113,
        "min": 3.130516767501831,
        "max": 3.130516767501831,
        "mean": 3.130516767501831,
        "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": 0.05714285714285714,
        "sum_squared": 0.0032653061224489793,
        "min": 0.05714285714285714,
        "max": 0.05714285714285714,
        "mean": 0.05714285714285714,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.04081632653061225,
        "sum_squared": 0.0016659725114535613,
        "min": 0.04081632653061225,
        "max": 0.04081632653061225,
        "mean": 0.04081632653061225,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.018518518518518517,
        "sum_squared": 0.0003429355281207133,
        "min": 0.018518518518518517,
        "max": 0.018518518518518517,
        "mean": 0.018518518518518517,
        "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": 59,
        "sum_squared": 3481,
        "min": 59,
        "max": 59,
        "mean": 59.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 293,
        "sum_squared": 85849,
        "min": 293,
        "max": 293,
        "mean": 293.0,
        "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": 68,
        "sum_squared": 4624,
        "min": 68,
        "max": 68,
        "mean": 68.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 68,
        "sum_squared": 4624,
        "min": 68,
        "max": 68,
        "mean": 68.0,
        "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": 3.316443681716919,
        "sum_squared": 10.998798694000072,
        "min": 3.316443681716919,
        "max": 3.316443681716919,
        "mean": 3.316443681716919,
        "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": 0.16949152542372883,
        "sum_squared": 0.02872737719046252,
        "min": 0.16949152542372883,
        "max": 0.16949152542372883,
        "mean": 0.16949152542372883,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0983606557377049,
        "sum_squared": 0.0096748185971513,
        "min": 0.0983606557377049,
        "max": 0.0983606557377049,
        "mean": 0.0983606557377049,
        "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": 68,
        "sum_squared": 4624,
        "min": 68,
        "max": 68,
        "mean": 68.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 302,
        "sum_squared": 91204,
        "min": 302,
        "max": 302,
        "mean": 302.0,
        "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": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "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": 2.8310747146606445,
        "sum_squared": 8.01498403999085,
        "min": 2.8310747146606445,
        "max": 2.8310747146606445,
        "mean": 2.8310747146606445,
        "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": 0.022222222222222227,
        "sum_squared": 0.0004938271604938274,
        "min": 0.022222222222222227,
        "max": 0.022222222222222227,
        "mean": 0.022222222222222227,
        "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": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 224,
        "sum_squared": 50176,
        "min": 224,
        "max": 224,
        "mean": 224.0,
        "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": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "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": 2.364560842514038,
        "sum_squared": 5.591147977950698,
        "min": 2.364560842514038,
        "max": 2.364560842514038,
        "mean": 2.364560842514038,
        "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": 0.047619047619047616,
        "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": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 215,
        "sum_squared": 46225,
        "min": 215,
        "max": 215,
        "mean": 215.0,
        "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": 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": 1.7652804851531982,
        "sum_squared": 3.116215191262711,
        "min": 1.7652804851531982,
        "max": 1.7652804851531982,
        "mean": 1.7652804851531982,
        "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": 0.03448275862068965,
        "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": 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": 147,
        "sum_squared": 21609,
        "min": 147,
        "max": 147,
        "mean": 147.0,
        "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": 30,
        "sum_squared": 900,
        "min": 30,
        "max": 30,
        "mean": 30.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 30,
        "sum_squared": 900,
        "min": 30,
        "max": 30,
        "mean": 30.0,
        "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": 2.065981864929199,
        "sum_squared": 4.268281066216332,
        "min": 2.065981864929199,
        "max": 2.065981864929199,
        "mean": 2.065981864929199,
        "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": 0.09523809523809523,
        "sum_squared": 0.009070294784580497,
        "min": 0.09523809523809523,
        "max": 0.09523809523809523,
        "mean": 0.09523809523809523,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.13793103448275862,
        "sum_squared": 0.019024970273483946,
        "min": 0.13793103448275862,
        "max": 0.13793103448275862,
        "mean": 0.13793103448275862,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.10714285714285714,
        "sum_squared": 0.011479591836734693,
        "min": 0.10714285714285714,
        "max": 0.10714285714285714,
        "mean": 0.10714285714285714,
        "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": 30,
        "sum_squared": 900,
        "min": 30,
        "max": 30,
        "mean": 30.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 156,
        "sum_squared": 24336,
        "min": 156,
        "max": 156,
        "mean": 156.0,
        "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": 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": 4.774997234344482,
        "sum_squared": 22.800598587997456,
        "min": 4.774997234344482,
        "max": 4.774997234344482,
        "mean": 4.774997234344482,
        "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": 0.018867924528301883,
        "sum_squared": 0.00035599857600569583,
        "min": 0.018867924528301883,
        "max": 0.018867924528301883,
        "mean": 0.018867924528301883,
        "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": 287,
        "sum_squared": 82369,
        "min": 287,
        "max": 287,
        "mean": 287.0,
        "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": 86,
        "sum_squared": 7396,
        "min": 86,
        "max": 86,
        "mean": 86.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 86,
        "sum_squared": 7396,
        "min": 86,
        "max": 86,
        "mean": 86.0,
        "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": 3.2750749588012695,
        "sum_squared": 10.726115985767137,
        "min": 3.2750749588012695,
        "max": 3.2750749588012695,
        "mean": 3.2750749588012695,
        "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": 0.028169014084507043,
        "sum_squared": 0.0007934933544931561,
        "min": 0.028169014084507043,
        "max": 0.028169014084507043,
        "mean": 0.028169014084507043,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.012658227848101266,
        "sum_squared": 0.0001602307322544464,
        "min": 0.012658227848101266,
        "max": 0.012658227848101266,
        "mean": 0.012658227848101266,
        "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": 86,
        "sum_squared": 7396,
        "min": 86,
        "max": 86,
        "mean": 86.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 423,
        "sum_squared": 178929,
        "min": 423,
        "max": 423,
        "mean": 423.0,
        "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": 45,
        "sum_squared": 2025,
        "min": 45,
        "max": 45,
        "mean": 45.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 45,
        "sum_squared": 2025,
        "min": 45,
        "max": 45,
        "mean": 45.0,
        "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": 3.137782096862793,
        "sum_squared": 9.845676487392666,
        "min": 3.137782096862793,
        "max": 3.137782096862793,
        "mean": 3.137782096862793,
        "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": 0.1875,
        "sum_squared": 0.03515625,
        "min": 0.1875,
        "max": 0.1875,
        "mean": 0.1875,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.1395348837209302,
        "sum_squared": 0.01946998377501351,
        "min": 0.1395348837209302,
        "max": 0.1395348837209302,
        "mean": 0.1395348837209302,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.07142857142857141,
        "sum_squared": 0.0051020408163265285,
        "min": 0.07142857142857141,
        "max": 0.07142857142857141,
        "mean": 0.07142857142857141,
        "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": 45,
        "sum_squared": 2025,
        "min": 45,
        "max": 45,
        "mean": 45.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 211,
        "sum_squared": 44521,
        "min": 211,
        "max": 211,
        "mean": 211.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  }
]