[
  {
    "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": 1735,
        "sum_squared": 3010225,
        "min": 1735,
        "max": 1735,
        "mean": 1735.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.652876615524292,
        "sum_squared": 0.4262478750984542,
        "min": 0.652876615524292,
        "max": 0.652876615524292,
        "mean": 0.652876615524292,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.8571428571428571,
        "sum_squared": 0.7346938775510203,
        "min": 0.8571428571428571,
        "max": 0.8571428571428571,
        "mean": 0.8571428571428571,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 20,
        "sum_squared": 400,
        "min": 20,
        "max": 20,
        "mean": 20.0,
        "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": 1598,
        "sum_squared": 2553604,
        "min": 1598,
        "max": 1598,
        "mean": 1598.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5993659496307373,
        "sum_squared": 0.35923954157675553,
        "min": 0.5993659496307373,
        "max": 0.5993659496307373,
        "mean": 0.5993659496307373,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "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": 1631,
        "sum_squared": 2660161,
        "min": 1631,
        "max": 1631,
        "mean": 1631.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6182413101196289,
        "sum_squared": 0.38222231753843516,
        "min": 0.6182413101196289,
        "max": 0.6182413101196289,
        "mean": 0.6182413101196289,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "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": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "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": 1717,
        "sum_squared": 2948089,
        "min": 1717,
        "max": 1717,
        "mean": 1717.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6172235012054443,
        "sum_squared": 0.38096485044030715,
        "min": 0.6172235012054443,
        "max": 0.6172235012054443,
        "mean": 0.6172235012054443,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "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": 1743,
        "sum_squared": 3038049,
        "min": 1743,
        "max": 1743,
        "mean": 1743.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.621185302734375,
        "sum_squared": 0.3858711803331971,
        "min": 0.621185302734375,
        "max": 0.621185302734375,
        "mean": 0.621185302734375,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5882352941176471,
        "sum_squared": 0.34602076124567477,
        "min": 0.5882352941176471,
        "max": 0.5882352941176471,
        "mean": 0.5882352941176471,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.4,
        "sum_squared": 0.16000000000000003,
        "min": 0.4,
        "max": 0.4,
        "mean": 0.4,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "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": 1851,
        "sum_squared": 3426201,
        "min": 1851,
        "max": 1851,
        "mean": 1851.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6454408168792725,
        "sum_squared": 0.4165938480937825,
        "min": 0.6454408168792725,
        "max": 0.6454408168792725,
        "mean": 0.6454408168792725,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5714285714285714,
        "sum_squared": 0.32653061224489793,
        "min": 0.5714285714285714,
        "max": 0.5714285714285714,
        "mean": 0.5714285714285714,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.375,
        "sum_squared": 0.140625,
        "min": 0.375,
        "max": 0.375,
        "mean": 0.375,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.30769230769230765,
        "sum_squared": 0.09467455621301772,
        "min": 0.30769230769230765,
        "max": 0.30769230769230765,
        "mean": 0.30769230769230765,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 71,
        "sum_squared": 5041,
        "min": 71,
        "max": 71,
        "mean": 71.0,
        "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": 1563,
        "sum_squared": 2442969,
        "min": 1563,
        "max": 1563,
        "mean": 1563.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5354666709899902,
        "sum_squared": 0.28672455574110245,
        "min": 0.5354666709899902,
        "max": 0.5354666709899902,
        "mean": 0.5354666709899902,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.4,
        "sum_squared": 0.16000000000000003,
        "min": 0.4,
        "max": 0.4,
        "mean": 0.4,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "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": 1893,
        "sum_squared": 3583449,
        "min": 1893,
        "max": 1893,
        "mean": 1893.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5199398994445801,
        "sum_squared": 0.27033749903444004,
        "min": 0.5199398994445801,
        "max": 0.5199398994445801,
        "mean": 0.5199398994445801,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 1641,
        "sum_squared": 2692881,
        "min": 1641,
        "max": 1641,
        "mean": 1641.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.672633171081543,
        "sum_squared": 0.45243538283921225,
        "min": 0.672633171081543,
        "max": 0.672633171081543,
        "mean": 0.672633171081543,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.7788007830714049,
        "sum_squared": 0.6065306597126334,
        "min": 0.7788007830714049,
        "max": 0.7788007830714049,
        "mean": 0.7788007830714049,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.7788007830714049,
        "sum_squared": 0.6065306597126334,
        "min": 0.7788007830714049,
        "max": 0.7788007830714049,
        "mean": 0.7788007830714049,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 28,
        "sum_squared": 784,
        "min": 28,
        "max": 28,
        "mean": 28.0,
        "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": 1935,
        "sum_squared": 3744225,
        "min": 1935,
        "max": 1935,
        "mean": 1935.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6182992458343506,
        "sum_squared": 0.3822939573993267,
        "min": 0.6182992458343506,
        "max": 0.6182992458343506,
        "mean": 0.6182992458343506,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 25,
        "sum_squared": 625,
        "min": 25,
        "max": 25,
        "mean": 25.0,
        "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": 1802,
        "sum_squared": 3247204,
        "min": 1802,
        "max": 1802,
        "mean": 1802.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6608259677886963,
        "sum_squared": 0.43669095970386707,
        "min": 0.6608259677886963,
        "max": 0.6608259677886963,
        "mean": 0.6608259677886963,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.23529411764705885,
        "sum_squared": 0.05536332179930797,
        "min": 0.23529411764705885,
        "max": 0.23529411764705885,
        "mean": 0.23529411764705885,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 64,
        "sum_squared": 4096,
        "min": 64,
        "max": 64,
        "mean": 64.0,
        "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": 1945,
        "sum_squared": 3783025,
        "min": 1945,
        "max": 1945,
        "mean": 1945.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6166229248046875,
        "sum_squared": 0.3802238313946873,
        "min": 0.6166229248046875,
        "max": 0.6166229248046875,
        "mean": 0.6166229248046875,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.36363636363636365,
        "sum_squared": 0.1322314049586777,
        "min": 0.36363636363636365,
        "max": 0.36363636363636365,
        "mean": 0.36363636363636365,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.30769230769230765,
        "sum_squared": 0.09467455621301772,
        "min": 0.30769230769230765,
        "max": 0.30769230769230765,
        "mean": 0.30769230769230765,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 37,
        "sum_squared": 1369,
        "min": 37,
        "max": 37,
        "mean": 37.0,
        "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": 1365,
        "sum_squared": 1863225,
        "min": 1365,
        "max": 1365,
        "mean": 1365.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.471085786819458,
        "sum_squared": 0.22192181854330784,
        "min": 0.471085786819458,
        "max": 0.471085786819458,
        "mean": 0.471085786819458,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.7499999999999999,
        "sum_squared": 0.5624999999999999,
        "min": 0.7499999999999999,
        "max": 0.7499999999999999,
        "mean": 0.7499999999999999,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.75,
        "sum_squared": 0.5625,
        "min": 0.75,
        "max": 0.75,
        "mean": 0.75,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.7,
        "sum_squared": 0.48999999999999994,
        "min": 0.7,
        "max": 0.7,
        "mean": 0.7,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5714285714285714,
        "sum_squared": 0.32653061224489793,
        "min": 0.5714285714285714,
        "max": 0.5714285714285714,
        "mean": 0.5714285714285714,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 44,
        "sum_squared": 1936,
        "min": 44,
        "max": 44,
        "mean": 44.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "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": 1779,
        "sum_squared": 3164841,
        "min": 1779,
        "max": 1779,
        "mean": 1779.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.373246431350708,
        "sum_squared": 0.13931289851603879,
        "min": 0.373246431350708,
        "max": 0.373246431350708,
        "mean": 0.373246431350708,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "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": 1524,
        "sum_squared": 2322576,
        "min": 1524,
        "max": 1524,
        "mean": 1524.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.386688232421875,
        "sum_squared": 0.14952778909355402,
        "min": 0.386688232421875,
        "max": 0.386688232421875,
        "mean": 0.386688232421875,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 1747,
        "sum_squared": 3052009,
        "min": 1747,
        "max": 1747,
        "mean": 1747.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.7278769016265869,
        "sum_squared": 0.5298047839215201,
        "min": 0.7278769016265869,
        "max": 0.7278769016265869,
        "mean": 0.7278769016265869,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.3333333333333333,
        "sum_squared": 0.1111111111111111,
        "min": 0.3333333333333333,
        "max": 0.3333333333333333,
        "mean": 0.3333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.375,
        "sum_squared": 0.140625,
        "min": 0.375,
        "max": 0.375,
        "mean": 0.375,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.13333333333333333,
        "sum_squared": 0.017777777777777778,
        "min": 0.13333333333333333,
        "max": 0.13333333333333333,
        "mean": 0.13333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 77,
        "sum_squared": 5929,
        "min": 77,
        "max": 77,
        "mean": 77.0,
        "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": 1499,
        "sum_squared": 2247001,
        "min": 1499,
        "max": 1499,
        "mean": 1499.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5581221580505371,
        "sum_squared": 0.3115003433069887,
        "min": 0.5581221580505371,
        "max": 0.5581221580505371,
        "mean": 0.5581221580505371,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.35294117647058826,
        "sum_squared": 0.12456747404844293,
        "min": 0.35294117647058826,
        "max": 0.35294117647058826,
        "mean": 0.35294117647058826,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3157894736842105,
        "sum_squared": 0.09972299168975068,
        "min": 0.3157894736842105,
        "max": 0.3157894736842105,
        "mean": 0.3157894736842105,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.36363636363636365,
        "sum_squared": 0.1322314049586777,
        "min": 0.36363636363636365,
        "max": 0.36363636363636365,
        "mean": 0.36363636363636365,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 60,
        "sum_squared": 3600,
        "min": 60,
        "max": 60,
        "mean": 60.0,
        "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": 1837,
        "sum_squared": 3374569,
        "min": 1837,
        "max": 1837,
        "mean": 1837.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.401228666305542,
        "sum_squared": 0.16098444266532397,
        "min": 0.401228666305542,
        "max": 0.401228666305542,
        "mean": 0.401228666305542,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.4,
        "sum_squared": 0.16000000000000003,
        "min": 0.4,
        "max": 0.4,
        "mean": 0.4,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.049787068367863944,
        "sum_squared": 0.0024787521766663585,
        "min": 0.049787068367863944,
        "max": 0.049787068367863944,
        "mean": 0.049787068367863944,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 1.10779904317048e-309,
        "sum_squared": 0.0,
        "min": 1.10779904317048e-309,
        "max": 1.10779904317048e-309,
        "mean": 1.10779904317048e-309,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 6,
        "sum_squared": 36,
        "min": 6,
        "max": 6,
        "mean": 6.0,
        "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": 1387,
        "sum_squared": 1923769,
        "min": 1387,
        "max": 1387,
        "mean": 1387.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.8469135761260986,
        "sum_squared": 0.7172626054266971,
        "min": 0.8469135761260986,
        "max": 0.8469135761260986,
        "mean": 0.8469135761260986,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666667,
        "sum_squared": 0.44444444444444453,
        "min": 0.6666666666666667,
        "max": 0.6666666666666667,
        "mean": 0.6666666666666667,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.47058823529411764,
        "sum_squared": 0.22145328719723184,
        "min": 0.47058823529411764,
        "max": 0.47058823529411764,
        "mean": 0.47058823529411764,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.14285714285714285,
        "sum_squared": 0.02040816326530612,
        "min": 0.14285714285714285,
        "max": 0.14285714285714285,
        "mean": 0.14285714285714285,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 76,
        "sum_squared": 5776,
        "min": 76,
        "max": 76,
        "mean": 76.0,
        "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": 1867,
        "sum_squared": 3485689,
        "min": 1867,
        "max": 1867,
        "mean": 1867.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5248615741729736,
        "sum_squared": 0.2754796720433319,
        "min": 0.5248615741729736,
        "max": 0.5248615741729736,
        "mean": 0.5248615741729736,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "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": 1673,
        "sum_squared": 2798929,
        "min": 1673,
        "max": 1673,
        "mean": 1673.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.4611694812774658,
        "sum_squared": 0.2126772904617269,
        "min": 0.4611694812774658,
        "max": 0.4611694812774658,
        "mean": 0.4611694812774658,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 30,
        "sum_squared": 900,
        "min": 30,
        "max": 30,
        "mean": 30.0,
        "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": 1889,
        "sum_squared": 3568321,
        "min": 1889,
        "max": 1889,
        "mean": 1889.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.4102494716644287,
        "sum_squared": 0.1683046290009429,
        "min": 0.4102494716644287,
        "max": 0.4102494716644287,
        "mean": 0.4102494716644287,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 9,
        "sum_squared": 81,
        "min": 9,
        "max": 9,
        "mean": 9.0,
        "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": 1617,
        "sum_squared": 2614689,
        "min": 1617,
        "max": 1617,
        "mean": 1617.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6484732627868652,
        "sum_squared": 0.4205175725494428,
        "min": 0.6484732627868652,
        "max": 0.6484732627868652,
        "mean": 0.6484732627868652,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "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": 1869,
        "sum_squared": 3493161,
        "min": 1869,
        "max": 1869,
        "mean": 1869.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5015168190002441,
        "sum_squared": 0.25151911974012364,
        "min": 0.5015168190002441,
        "max": 0.5015168190002441,
        "mean": 0.5015168190002441,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 1768,
        "sum_squared": 3125824,
        "min": 1768,
        "max": 1768,
        "mean": 1768.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6585769653320312,
        "sum_squared": 0.4337236192659475,
        "min": 0.6585769653320312,
        "max": 0.6585769653320312,
        "mean": 0.6585769653320312,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.2222222222222222,
        "sum_squared": 0.04938271604938271,
        "min": 0.2222222222222222,
        "max": 0.2222222222222222,
        "mean": 0.2222222222222222,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 28,
        "sum_squared": 784,
        "min": 28,
        "max": 28,
        "mean": 28.0,
        "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": 1818,
        "sum_squared": 3305124,
        "min": 1818,
        "max": 1818,
        "mean": 1818.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5825033187866211,
        "sum_squared": 0.3393101163974279,
        "min": 0.5825033187866211,
        "max": 0.5825033187866211,
        "mean": 0.5825033187866211,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "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": 1724,
        "sum_squared": 2972176,
        "min": 1724,
        "max": 1724,
        "mean": 1724.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.3173825740814209,
        "sum_squared": 0.10073169833054862,
        "min": 0.3173825740814209,
        "max": 0.3173825740814209,
        "mean": 0.3173825740814209,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "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": 1397,
        "sum_squared": 1951609,
        "min": 1397,
        "max": 1397,
        "mean": 1397.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.8020062446594238,
        "sum_squared": 0.6432140164727116,
        "min": 0.8020062446594238,
        "max": 0.8020062446594238,
        "mean": 0.8020062446594238,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.8461538461538461,
        "sum_squared": 0.7159763313609467,
        "min": 0.8461538461538461,
        "max": 0.8461538461538461,
        "mean": 0.8461538461538461,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.896551724137931,
        "sum_squared": 0.8038049940546969,
        "min": 0.896551724137931,
        "max": 0.896551724137931,
        "mean": 0.896551724137931,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.8125,
        "sum_squared": 0.66015625,
        "min": 0.8125,
        "max": 0.8125,
        "mean": 0.8125,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.38461538461538464,
        "sum_squared": 0.14792899408284024,
        "min": 0.38461538461538464,
        "max": 0.38461538461538464,
        "mean": 0.38461538461538464,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 88,
        "sum_squared": 7744,
        "min": 88,
        "max": 88,
        "mean": 88.0,
        "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": 1412,
        "sum_squared": 1993744,
        "min": 1412,
        "max": 1412,
        "mean": 1412.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6000654697418213,
        "sum_squared": 0.36007856797647264,
        "min": 0.6000654697418213,
        "max": 0.6000654697418213,
        "mean": 0.6000654697418213,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.3333333333333333,
        "sum_squared": 0.1111111111111111,
        "min": 0.3333333333333333,
        "max": 0.3333333333333333,
        "mean": 0.3333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 39,
        "sum_squared": 1521,
        "min": 39,
        "max": 39,
        "mean": 39.0,
        "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": 1367,
        "sum_squared": 1868689,
        "min": 1367,
        "max": 1367,
        "mean": 1367.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6082258224487305,
        "sum_squared": 0.3699386510934346,
        "min": 0.6082258224487305,
        "max": 0.6082258224487305,
        "mean": 0.6082258224487305,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.16666666666666669,
        "sum_squared": 0.027777777777777783,
        "min": 0.16666666666666669,
        "max": 0.16666666666666669,
        "mean": 0.16666666666666669,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 29,
        "sum_squared": 841,
        "min": 29,
        "max": 29,
        "mean": 29.0,
        "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": 1842,
        "sum_squared": 3392964,
        "min": 1842,
        "max": 1842,
        "mean": 1842.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5674817562103271,
        "sum_squared": 0.3220355436315572,
        "min": 0.5674817562103271,
        "max": 0.5674817562103271,
        "mean": 0.5674817562103271,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.5882352941176471,
        "sum_squared": 0.34602076124567477,
        "min": 0.5882352941176471,
        "max": 0.5882352941176471,
        "mean": 0.5882352941176471,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5714285714285715,
        "sum_squared": 0.32653061224489804,
        "min": 0.5714285714285715,
        "max": 0.5714285714285715,
        "mean": 0.5714285714285715,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.375,
        "sum_squared": 0.140625,
        "min": 0.375,
        "max": 0.375,
        "mean": 0.375,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 77,
        "sum_squared": 5929,
        "min": 77,
        "max": 77,
        "mean": 77.0,
        "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": 1843,
        "sum_squared": 3396649,
        "min": 1843,
        "max": 1843,
        "mean": 1843.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.4308292865753174,
        "sum_squared": 0.18561387417099695,
        "min": 0.4308292865753174,
        "max": 0.4308292865753174,
        "mean": 0.4308292865753174,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.36787944117144233,
        "sum_squared": 0.1353352832366127,
        "min": 0.36787944117144233,
        "max": 0.36787944117144233,
        "mean": 0.36787944117144233,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 8.18558927632837e-309,
        "sum_squared": 0.0,
        "min": 8.18558927632837e-309,
        "max": 8.18558927632837e-309,
        "mean": 8.18558927632837e-309,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "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": 1865,
        "sum_squared": 3478225,
        "min": 1865,
        "max": 1865,
        "mean": 1865.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6287984848022461,
        "sum_squared": 0.3953875344896005,
        "min": 0.6287984848022461,
        "max": 0.6287984848022461,
        "mean": 0.6287984848022461,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "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": 1780,
        "sum_squared": 3168400,
        "min": 1780,
        "max": 1780,
        "mean": 1780.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.7065272331237793,
        "sum_squared": 0.4991807311455432,
        "min": 0.7065272331237793,
        "max": 0.7065272331237793,
        "mean": 0.7065272331237793,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.8571428571428571,
        "sum_squared": 0.7346938775510203,
        "min": 0.8571428571428571,
        "max": 0.8571428571428571,
        "mean": 0.8571428571428571,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.7142857142857143,
        "sum_squared": 0.5102040816326531,
        "min": 0.7142857142857143,
        "max": 0.7142857142857143,
        "mean": 0.7142857142857143,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 39,
        "sum_squared": 1521,
        "min": 39,
        "max": 39,
        "mean": 39.0,
        "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": 1813,
        "sum_squared": 3286969,
        "min": 1813,
        "max": 1813,
        "mean": 1813.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5104985237121582,
        "sum_squared": 0.26060874271229295,
        "min": 0.5104985237121582,
        "max": 0.5104985237121582,
        "mean": 0.5104985237121582,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.36787944117144233,
        "sum_squared": 0.1353352832366127,
        "min": 0.36787944117144233,
        "max": 0.36787944117144233,
        "mean": 0.36787944117144233,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 8.18558927632837e-309,
        "sum_squared": 0.0,
        "min": 8.18558927632837e-309,
        "max": 8.18558927632837e-309,
        "mean": 8.18558927632837e-309,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 1908,
        "sum_squared": 3640464,
        "min": 1908,
        "max": 1908,
        "mean": 1908.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.3500406742095947,
        "sum_squared": 0.12252847360110763,
        "min": 0.3500406742095947,
        "max": 0.3500406742095947,
        "mean": 0.3500406742095947,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "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": 1806,
        "sum_squared": 3261636,
        "min": 1806,
        "max": 1806,
        "mean": 1806.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.620819091796875,
        "sum_squared": 0.3854163447394967,
        "min": 0.620819091796875,
        "max": 0.620819091796875,
        "mean": 0.620819091796875,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.7272727272727273,
        "sum_squared": 0.5289256198347108,
        "min": 0.7272727272727273,
        "max": 0.7272727272727273,
        "mean": 0.7272727272727273,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.7272727272727272,
        "sum_squared": 0.5289256198347106,
        "min": 0.7272727272727272,
        "max": 0.7272727272727272,
        "mean": 0.7272727272727272,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5362560368285115,
        "sum_squared": 0.2875705370350219,
        "min": 0.5362560368285115,
        "max": 0.5362560368285115,
        "mean": 0.5362560368285115,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.33516002301781966,
        "sum_squared": 0.1123322410293054,
        "min": 0.33516002301781966,
        "max": 0.33516002301781966,
        "mean": 0.33516002301781966,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 31,
        "sum_squared": 961,
        "min": 31,
        "max": 31,
        "mean": 31.0,
        "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": 1828,
        "sum_squared": 3341584,
        "min": 1828,
        "max": 1828,
        "mean": 1828.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6025104522705078,
        "sum_squared": 0.3630188450952119,
        "min": 0.6025104522705078,
        "max": 0.6025104522705078,
        "mean": 0.6025104522705078,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "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": 1708,
        "sum_squared": 2917264,
        "min": 1708,
        "max": 1708,
        "mean": 1708.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.4029567241668701,
        "sum_squared": 0.16237412155129505,
        "min": 0.4029567241668701,
        "max": 0.4029567241668701,
        "mean": 0.4029567241668701,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "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": 1491,
        "sum_squared": 2223081,
        "min": 1491,
        "max": 1491,
        "mean": 1491.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.4398353099822998,
        "sum_squared": 0.19345509990722576,
        "min": 0.4398353099822998,
        "max": 0.4398353099822998,
        "mean": 0.4398353099822998,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.5454545454545454,
        "sum_squared": 0.29752066115702475,
        "min": 0.5454545454545454,
        "max": 0.5454545454545454,
        "mean": 0.5454545454545454,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5454545454545454,
        "sum_squared": 0.29752066115702475,
        "min": 0.5454545454545454,
        "max": 0.5454545454545454,
        "mean": 0.5454545454545454,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3333333333333333,
        "sum_squared": 0.1111111111111111,
        "min": 0.3333333333333333,
        "max": 0.3333333333333333,
        "mean": 0.3333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 36,
        "sum_squared": 1296,
        "min": 36,
        "max": 36,
        "mean": 36.0,
        "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": 1894,
        "sum_squared": 3587236,
        "min": 1894,
        "max": 1894,
        "mean": 1894.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6962099075317383,
        "sum_squared": 0.48470823534535157,
        "min": 0.6962099075317383,
        "max": 0.6962099075317383,
        "mean": 0.6962099075317383,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 1719,
        "sum_squared": 2954961,
        "min": 1719,
        "max": 1719,
        "mean": 1719.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5318849086761475,
        "sum_squared": 0.2829015560774337,
        "min": 0.5318849086761475,
        "max": 0.5318849086761475,
        "mean": 0.5318849086761475,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "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": 1870,
        "sum_squared": 3496900,
        "min": 1870,
        "max": 1870,
        "mean": 1870.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6119248867034912,
        "sum_squared": 0.37445206696708055,
        "min": 0.6119248867034912,
        "max": 0.6119248867034912,
        "mean": 0.6119248867034912,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.4,
        "sum_squared": 0.16000000000000003,
        "min": 0.4,
        "max": 0.4,
        "mean": 0.4,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 1809,
        "sum_squared": 3272481,
        "min": 1809,
        "max": 1809,
        "mean": 1809.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6357569694519043,
        "sum_squared": 0.4041869242066696,
        "min": 0.6357569694519043,
        "max": 0.6357569694519043,
        "mean": 0.6357569694519043,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "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": 1796,
        "sum_squared": 3225616,
        "min": 1796,
        "max": 1796,
        "mean": 1796.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.5794453620910645,
        "sum_squared": 0.3357569276488448,
        "min": 0.5794453620910645,
        "max": 0.5794453620910645,
        "mean": 0.5794453620910645,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.2857142857142857,
        "sum_squared": 0.08163265306122448,
        "min": 0.2857142857142857,
        "max": 0.2857142857142857,
        "mean": 0.2857142857142857,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0009118819655545162,
        "sum_squared": 8.31528719103568e-07,
        "min": 0.0009118819655545162,
        "max": 0.0009118819655545162,
        "mean": 0.0009118819655545162,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.0290047235994e-311,
        "sum_squared": 0.0,
        "min": 2.0290047235994e-311,
        "max": 2.0290047235994e-311,
        "mean": 2.0290047235994e-311,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 1767,
        "sum_squared": 3122289,
        "min": 1767,
        "max": 1767,
        "mean": 1767.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.3251361846923828,
        "sum_squared": 0.10571353859631927,
        "min": 0.3251361846923828,
        "max": 0.3251361846923828,
        "mean": 0.3251361846923828,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 1708,
        "sum_squared": 2917264,
        "min": 1708,
        "max": 1708,
        "mean": 1708.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.48076844215393066,
        "sum_squared": 0.23113829497111738,
        "min": 0.48076844215393066,
        "max": 0.48076844215393066,
        "mean": 0.48076844215393066,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "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": "rouge_l",
          "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": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.36787944117144233,
        "sum_squared": 0.1353352832366127,
        "min": 0.36787944117144233,
        "max": 0.36787944117144233,
        "mean": 0.36787944117144233,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 8.18558927632837e-309,
        "sum_squared": 0.0,
        "min": 8.18558927632837e-309,
        "max": 8.18558927632837e-309,
        "mean": 8.18558927632837e-309,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 1603,
        "sum_squared": 2569609,
        "min": 1603,
        "max": 1603,
        "mean": 1603.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.4881293773651123,
        "sum_squared": 0.2382702890468522,
        "min": 0.4881293773651123,
        "max": 0.4881293773651123,
        "mean": 0.4881293773651123,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "sum_squared": 0.4444444444444444,
        "min": 0.6666666666666666,
        "max": 0.6666666666666666,
        "mean": 0.6666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "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": 1456,
        "sum_squared": 2119936,
        "min": 1456,
        "max": 1456,
        "mean": 1456.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.325824499130249,
        "sum_squared": 0.10616160423347765,
        "min": 0.325824499130249,
        "max": 0.325824499130249,
        "mean": 0.325824499130249,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "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": 1831,
        "sum_squared": 3352561,
        "min": 1831,
        "max": 1831,
        "mean": 1831.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "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": 0.41121459007263184,
        "sum_squared": 0.16909743908860264,
        "min": 0.41121459007263184,
        "max": 0.41121459007263184,
        "mean": 0.41121459007263184,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.16666666666666666,
        "sum_squared": 0.027777777777777776,
        "min": 0.16666666666666666,
        "max": 0.16666666666666666,
        "mean": 0.16666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3076923076923077,
        "sum_squared": 0.09467455621301776,
        "min": 0.3076923076923077,
        "max": 0.3076923076923077,
        "mean": 0.3076923076923077,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.11111111111111109,
        "sum_squared": 0.012345679012345675,
        "min": 0.11111111111111109,
        "max": 0.11111111111111109,
        "mean": 0.11111111111111109,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 52,
        "sum_squared": 2704,
        "min": 52,
        "max": 52,
        "mean": 52.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  }
]