[
  {
    "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": 3511,
        "sum_squared": 12327121,
        "min": 3511,
        "max": 3511,
        "mean": 3511.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "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": 0.9186711311340332,
        "sum_squared": 0.843956647179084,
        "min": 0.9186711311340332,
        "max": 0.9186711311340332,
        "mean": 0.9186711311340332,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 0.4,
        "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": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "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": 3375,
        "sum_squared": 11390625,
        "min": 3375,
        "max": 3375,
        "mean": 3375.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.4032626152038574,
        "sum_squared": 0.16262073682105438,
        "min": 0.4032626152038574,
        "max": 0.4032626152038574,
        "mean": 0.4032626152038574,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "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,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 3399,
        "sum_squared": 11553201,
        "min": 3399,
        "max": 3399,
        "mean": 3399.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.889366865158081,
        "sum_squared": 0.7909734208411123,
        "min": 0.889366865158081,
        "max": 0.889366865158081,
        "mean": 0.889366865158081,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 9,
        "sum_squared": 81,
        "min": 9,
        "max": 9,
        "mean": 9.0,
        "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": 3496,
        "sum_squared": 12222016,
        "min": 3496,
        "max": 3496,
        "mean": 3496.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.9155101776123047,
        "sum_squared": 0.8381588853117137,
        "min": 0.9155101776123047,
        "max": 0.9155101776123047,
        "mean": 0.9155101776123047,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 11,
        "sum_squared": 121,
        "min": 11,
        "max": 11,
        "mean": 11.0,
        "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": 3042,
        "sum_squared": 9253764,
        "min": 3042,
        "max": 3042,
        "mean": 3042.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 0.2613513469696045,
        "sum_squared": 0.0683045265628266,
        "min": 0.2613513469696045,
        "max": 0.2613513469696045,
        "mean": 0.2613513469696045,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 0.47768754038252614,
        "sum_squared": 0.22818538623670753,
        "min": 0.47768754038252614,
        "max": 0.47768754038252614,
        "mean": 0.47768754038252614,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 1.594335087959687e-308,
        "sum_squared": 0.0,
        "min": 1.594335087959687e-308,
        "max": 1.594335087959687e-308,
        "mean": 1.594335087959687e-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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 18,
        "sum_squared": 324,
        "min": 18,
        "max": 18,
        "mean": 18.0,
        "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": 3630,
        "sum_squared": 13176900,
        "min": 3630,
        "max": 3630,
        "mean": 3630.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0612504482269287,
        "sum_squared": 1.126252513861857,
        "min": 1.0612504482269287,
        "max": 1.0612504482269287,
        "mean": 1.0612504482269287,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.14285714285714285,
        "sum_squared": 0.02040816326530612,
        "min": 0.14285714285714285,
        "max": 0.14285714285714285,
        "mean": 0.14285714285714285,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.14285714285714288,
        "sum_squared": 0.020408163265306128,
        "min": 0.14285714285714288,
        "max": 0.14285714285714288,
        "mean": 0.14285714285714288,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0909090909090909,
        "sum_squared": 0.008264462809917354,
        "min": 0.0909090909090909,
        "max": 0.0909090909090909,
        "mean": 0.0909090909090909,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "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": 3321,
        "sum_squared": 11029041,
        "min": 3321,
        "max": 3321,
        "mean": 3321.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.835554838180542,
        "sum_squared": 0.6981518876069117,
        "min": 0.835554838180542,
        "max": 0.835554838180542,
        "mean": 0.835554838180542,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.4,
        "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": 0.3333333333333333,
        "sum_squared": 0.1111111111111111,
        "min": 0.3333333333333333,
        "max": 0.3333333333333333,
        "mean": 0.3333333333333333,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": "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": 3662,
        "sum_squared": 13410244,
        "min": 3662,
        "max": 3662,
        "mean": 3662.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.25126194953918457,
        "sum_squared": 0.06313256728623173,
        "min": 0.25126194953918457,
        "max": 0.25126194953918457,
        "mean": 0.25126194953918457,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": "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": 3428,
        "sum_squared": 11751184,
        "min": 3428,
        "max": 3428,
        "mean": 3428.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "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": 1.0240693092346191,
        "sum_squared": 1.04871795011627,
        "min": 1.0240693092346191,
        "max": 1.0240693092346191,
        "mean": 1.0240693092346191,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 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": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "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": 3228,
        "sum_squared": 10419984,
        "min": 3228,
        "max": 3228,
        "mean": 3228.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 11,
        "sum_squared": 121,
        "min": 11,
        "max": 11,
        "mean": 11.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 11,
        "sum_squared": 121,
        "min": 11,
        "max": 11,
        "mean": 11.0,
        "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": 0.9479615688323975,
        "sum_squared": 0.8986311359831802,
        "min": 0.9479615688323975,
        "max": 0.9479615688323975,
        "mean": 0.9479615688323975,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2,
        "sum_squared": 0.04000000000000001,
        "min": 0.2,
        "max": 0.2,
        "mean": 0.2,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.23529411764705882,
        "sum_squared": 0.05536332179930796,
        "min": 0.23529411764705882,
        "max": 0.23529411764705882,
        "mean": 0.23529411764705882,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2,
        "sum_squared": 0.04000000000000001,
        "min": 0.2,
        "max": 0.2,
        "mean": 0.2,
        "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": 11,
        "sum_squared": 121,
        "min": 11,
        "max": 11,
        "mean": 11.0,
        "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": "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": 3561,
        "sum_squared": 12680721,
        "min": 3561,
        "max": 3561,
        "mean": 3561.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0144038200378418,
        "sum_squared": 1.0290151101073661,
        "min": 1.0144038200378418,
        "max": 1.0144038200378418,
        "mean": 1.0144038200378418,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.7692307692307692,
        "sum_squared": 0.5917159763313609,
        "min": 0.7692307692307692,
        "max": 0.7692307692307692,
        "mean": 0.7692307692307692,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "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": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6363636363636364,
        "sum_squared": 0.4049586776859504,
        "min": 0.6363636363636364,
        "max": 0.6363636363636364,
        "mean": 0.6363636363636364,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "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": "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": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "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": 3706,
        "sum_squared": 13734436,
        "min": 3706,
        "max": 3706,
        "mean": 3706.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 0.8954699039459229,
        "sum_squared": 0.8018663488729203,
        "min": 0.8954699039459229,
        "max": 0.8954699039459229,
        "mean": 0.8954699039459229,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3333333333333333,
        "sum_squared": 0.1111111111111111,
        "min": 0.3333333333333333,
        "max": 0.3333333333333333,
        "mean": 0.3333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.26666666666666666,
        "sum_squared": 0.07111111111111111,
        "min": 0.26666666666666666,
        "max": 0.26666666666666666,
        "mean": 0.26666666666666666,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2,
        "sum_squared": 0.04000000000000001,
        "min": 0.2,
        "max": 0.2,
        "mean": 0.2,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 43,
        "sum_squared": 1849,
        "min": 43,
        "max": 43,
        "mean": 43.0,
        "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": 3755,
        "sum_squared": 14100025,
        "min": 3755,
        "max": 3755,
        "mean": 3755.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 11,
        "sum_squared": 121,
        "min": 11,
        "max": 11,
        "mean": 11.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 11,
        "sum_squared": 121,
        "min": 11,
        "max": 11,
        "mean": 11.0,
        "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": 0.3721733093261719,
        "sum_squared": 0.1385129721747944,
        "min": 0.3721733093261719,
        "max": 0.3721733093261719,
        "mean": 0.3721733093261719,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.8235294117647058,
        "sum_squared": 0.6782006920415224,
        "min": 0.8235294117647058,
        "max": 0.8235294117647058,
        "mean": 0.8235294117647058,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.8235294117647058,
        "sum_squared": 0.6782006920415224,
        "min": 0.8235294117647058,
        "max": 0.8235294117647058,
        "mean": 0.8235294117647058,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "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": "bleu_4",
          "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": "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": 11,
        "sum_squared": 121,
        "min": 11,
        "max": 11,
        "mean": 11.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 51,
        "sum_squared": 2601,
        "min": 51,
        "max": 51,
        "mean": 51.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "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": 3557,
        "sum_squared": 12652249,
        "min": 3557,
        "max": 3557,
        "mean": 3557.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 0.258669376373291,
        "sum_squared": 0.06690984627334728,
        "min": 0.258669376373291,
        "max": 0.258669376373291,
        "mean": 0.258669376373291,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.4,
        "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": 0.5714285714285715,
        "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": 0.47768754038252614,
        "sum_squared": 0.22818538623670753,
        "min": 0.47768754038252614,
        "max": 0.47768754038252614,
        "mean": 0.47768754038252614,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 1.594335087959687e-308,
        "sum_squared": 0.0,
        "min": 1.594335087959687e-308,
        "max": 1.594335087959687e-308,
        "mean": 1.594335087959687e-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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": "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": 3308,
        "sum_squared": 10942864,
        "min": 3308,
        "max": 3308,
        "mean": 3308.0,
        "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": 0.2193920612335205,
        "sum_squared": 0.04813287653229281,
        "min": 0.2193920612335205,
        "max": 0.2193920612335205,
        "mean": 0.2193920612335205,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": "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": 3520,
        "sum_squared": 12390400,
        "min": 3520,
        "max": 3520,
        "mean": 3520.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.2627694606781006,
        "sum_squared": 0.06904778946505985,
        "min": 0.2627694606781006,
        "max": 0.2627694606781006,
        "mean": 0.2627694606781006,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 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": 0.8,
        "sum_squared": 0.6400000000000001,
        "min": 0.8,
        "max": 0.8,
        "mean": 0.8,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.3032653298563167,
        "sum_squared": 0.09196986029286058,
        "min": 0.3032653298563167,
        "max": 0.3032653298563167,
        "mean": 0.3032653298563167,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.3495755153097447e-308,
        "sum_squared": 0.0,
        "min": 1.3495755153097447e-308,
        "max": 1.3495755153097447e-308,
        "mean": 1.3495755153097447e-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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 14,
        "sum_squared": 196,
        "min": 14,
        "max": 14,
        "mean": 14.0,
        "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": 3277,
        "sum_squared": 10738729,
        "min": 3277,
        "max": 3277,
        "mean": 3277.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 14,
        "sum_squared": 196,
        "min": 14,
        "max": 14,
        "mean": 14.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 14,
        "sum_squared": 196,
        "min": 14,
        "max": 14,
        "mean": 14.0,
        "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": 0.9844205379486084,
        "sum_squared": 0.9690837955350275,
        "min": 0.9844205379486084,
        "max": 0.9844205379486084,
        "mean": 0.9844205379486084,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3333333333333333,
        "sum_squared": 0.1111111111111111,
        "min": 0.3333333333333333,
        "max": 0.3333333333333333,
        "mean": 0.3333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.28571428571428575,
        "sum_squared": 0.08163265306122451,
        "min": 0.28571428571428575,
        "max": 0.28571428571428575,
        "mean": 0.28571428571428575,
        "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": 14,
        "sum_squared": 196,
        "min": 14,
        "max": 14,
        "mean": 14.0,
        "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": "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": 3150,
        "sum_squared": 9922500,
        "min": 3150,
        "max": 3150,
        "mean": 3150.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 9,
        "sum_squared": 81,
        "min": 9,
        "max": 9,
        "mean": 9.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 9,
        "sum_squared": 81,
        "min": 9,
        "max": 9,
        "mean": 9.0,
        "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": 0.6034114360809326,
        "sum_squared": 0.36410536119325343,
        "min": 0.6034114360809326,
        "max": 0.6034114360809326,
        "mean": 0.6034114360809326,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "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": 9,
        "sum_squared": 81,
        "min": 9,
        "max": 9,
        "mean": 9.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "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": 3772,
        "sum_squared": 14227984,
        "min": 3772,
        "max": 3772,
        "mean": 3772.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "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": 0.409564733505249,
        "sum_squared": 0.16774327093122565,
        "min": 0.409564733505249,
        "max": 0.409564733505249,
        "mean": 0.409564733505249,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.7058823529411764,
        "sum_squared": 0.49826989619377154,
        "min": 0.7058823529411764,
        "max": 0.7058823529411764,
        "mean": 0.7058823529411764,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.7000000000000001,
        "sum_squared": 0.4900000000000001,
        "min": 0.7000000000000001,
        "max": 0.7000000000000001,
        "mean": 0.7000000000000001,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6153846153846154,
        "sum_squared": 0.37869822485207105,
        "min": 0.6153846153846154,
        "max": 0.6153846153846154,
        "mean": 0.6153846153846154,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2,
        "sum_squared": 0.04000000000000001,
        "min": 0.2,
        "max": 0.2,
        "mean": 0.2,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "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": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 67,
        "sum_squared": 4489,
        "min": 67,
        "max": 67,
        "mean": 67.0,
        "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": 3643,
        "sum_squared": 13271449,
        "min": 3643,
        "max": 3643,
        "mean": 3643.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 0.2699699401855469,
        "sum_squared": 0.07288376860378776,
        "min": 0.2699699401855469,
        "max": 0.2699699401855469,
        "mean": 0.2699699401855469,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": "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": 3453,
        "sum_squared": 11923209,
        "min": 3453,
        "max": 3453,
        "mean": 3453.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 0.8493340015411377,
        "sum_squared": 0.7213682461738813,
        "min": 0.8493340015411377,
        "max": 0.8493340015411377,
        "mean": 0.8493340015411377,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": "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": 3677,
        "sum_squared": 13520329,
        "min": 3677,
        "max": 3677,
        "mean": 3677.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.8066091537475586,
        "sum_squared": 0.6506183269093526,
        "min": 0.8066091537475586,
        "max": 0.8066091537475586,
        "mean": 0.8066091537475586,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": "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": 3390,
        "sum_squared": 11492100,
        "min": 3390,
        "max": 3390,
        "mean": 3390.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 0.326526403427124,
        "sum_squared": 0.10661949213505295,
        "min": 0.326526403427124,
        "max": 0.326526403427124,
        "mean": 0.326526403427124,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 14,
        "sum_squared": 196,
        "min": 14,
        "max": 14,
        "mean": 14.0,
        "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": 3165,
        "sum_squared": 10017225,
        "min": 3165,
        "max": 3165,
        "mean": 3165.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.23354363441467285,
        "sum_squared": 0.054542629175614366,
        "min": 0.23354363441467285,
        "max": 0.23354363441467285,
        "mean": 0.23354363441467285,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "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": 3539,
        "sum_squared": 12524521,
        "min": 3539,
        "max": 3539,
        "mean": 3539.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "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": 0.8684811592102051,
        "sum_squared": 0.7542595239031016,
        "min": 0.8684811592102051,
        "max": 0.8684811592102051,
        "mean": 0.8684811592102051,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.5,
        "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": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 27,
        "sum_squared": 729,
        "min": 27,
        "max": 27,
        "mean": 27.0,
        "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": 3100,
        "sum_squared": 9610000,
        "min": 3100,
        "max": 3100,
        "mean": 3100.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "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": 0.8762445449829102,
        "sum_squared": 0.7678045026123073,
        "min": 0.8762445449829102,
        "max": 0.8762445449829102,
        "mean": 0.8762445449829102,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 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": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 18,
        "sum_squared": 324,
        "min": 18,
        "max": 18,
        "mean": 18.0,
        "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": 3494,
        "sum_squared": 12208036,
        "min": 3494,
        "max": 3494,
        "mean": 3494.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 0.5429232120513916,
        "sum_squared": 0.29476561418420033,
        "min": 0.5429232120513916,
        "max": 0.5429232120513916,
        "mean": 0.5429232120513916,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 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,
        "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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "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": 3783,
        "sum_squared": 14311089,
        "min": 3783,
        "max": 3783,
        "mean": 3783.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 18,
        "sum_squared": 324,
        "min": 18,
        "max": 18,
        "mean": 18.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 18,
        "sum_squared": 324,
        "min": 18,
        "max": 18,
        "mean": 18.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0977084636688232,
        "sum_squared": 1.2049638712101682,
        "min": 1.0977084636688232,
        "max": 1.0977084636688232,
        "mean": 1.0977084636688232,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.8461538461538461,
        "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": 18,
        "sum_squared": 324,
        "min": 18,
        "max": 18,
        "mean": 18.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 87,
        "sum_squared": 7569,
        "min": 87,
        "max": 87,
        "mean": 87.0,
        "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": 3820,
        "sum_squared": 14592400,
        "min": 3820,
        "max": 3820,
        "mean": 3820.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 0.44613194465637207,
        "sum_squared": 0.19903371204287623,
        "min": 0.44613194465637207,
        "max": 0.44613194465637207,
        "mean": 0.44613194465637207,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.5,
        "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": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 38,
        "sum_squared": 1444,
        "min": 38,
        "max": 38,
        "mean": 38.0,
        "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": 3755,
        "sum_squared": 14100025,
        "min": 3755,
        "max": 3755,
        "mean": 3755.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "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": 0.39412593841552734,
        "sum_squared": 0.15533525533192005,
        "min": 0.39412593841552734,
        "max": 0.39412593841552734,
        "mean": 0.39412593841552734,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.14285714285714285,
        "sum_squared": 0.02040816326530612,
        "min": 0.14285714285714285,
        "max": 0.14285714285714285,
        "mean": 0.14285714285714285,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "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": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.12500000000000003,
        "sum_squared": 0.015625000000000007,
        "min": 0.12500000000000003,
        "max": 0.12500000000000003,
        "mean": 0.12500000000000003,
        "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": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 33,
        "sum_squared": 1089,
        "min": 33,
        "max": 33,
        "mean": 33.0,
        "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": 3616,
        "sum_squared": 13075456,
        "min": 3616,
        "max": 3616,
        "mean": 3616.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "test"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0970735549926758,
        "sum_squared": 1.2035703850642676,
        "min": 1.0970735549926758,
        "max": 1.0970735549926758,
        "mean": 1.0970735549926758,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5882352941176471,
        "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": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 76,
        "sum_squared": 5776,
        "min": 76,
        "max": 76,
        "mean": 76.0,
        "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": 3590,
        "sum_squared": 12888100,
        "min": 3590,
        "max": 3590,
        "mean": 3590.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 6,
        "sum_squared": 36,
        "min": 6,
        "max": 6,
        "mean": 6.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 6,
        "sum_squared": 36,
        "min": 6,
        "max": 6,
        "mean": 6.0,
        "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": 0.9080181121826172,
        "sum_squared": 0.824496892051684,
        "min": 0.9080181121826172,
        "max": 0.9080181121826172,
        "mean": 0.9080181121826172,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.5454545454545455,
        "sum_squared": 0.29752066115702486,
        "min": 0.5454545454545455,
        "max": 0.5454545454545455,
        "mean": 0.5454545454545455,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.6153846153846153,
        "sum_squared": 0.3786982248520709,
        "min": 0.6153846153846153,
        "max": 0.6153846153846153,
        "mean": 0.6153846153846153,
        "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": 0.18393972058572117,
        "sum_squared": 0.033833820809153176,
        "min": 0.18393972058572117,
        "max": 0.18393972058572117,
        "mean": 0.18393972058572117,
        "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": 6,
        "sum_squared": 36,
        "min": 6,
        "max": 6,
        "mean": 6.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 24,
        "sum_squared": 576,
        "min": 24,
        "max": 24,
        "mean": 24.0,
        "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": 3637,
        "sum_squared": 13227769,
        "min": 3637,
        "max": 3637,
        "mean": 3637.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 0.8191604614257812,
        "sum_squared": 0.6710238615632989,
        "min": 0.8191604614257812,
        "max": 0.8191604614257812,
        "mean": 0.8191604614257812,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 14,
        "sum_squared": 196,
        "min": 14,
        "max": 14,
        "mean": 14.0,
        "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": 3563,
        "sum_squared": 12694969,
        "min": 3563,
        "max": 3563,
        "mean": 3563.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "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": 0.9107813835144043,
        "sum_squared": 0.8295227285564124,
        "min": 0.9107813835144043,
        "max": 0.9107813835144043,
        "mean": 0.9107813835144043,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.7272727272727273,
        "sum_squared": 0.5289256198347108,
        "min": 0.7272727272727273,
        "max": 0.7272727272727273,
        "mean": 0.7272727272727273,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.923076923076923,
        "sum_squared": 0.8520710059171597,
        "min": 0.923076923076923,
        "max": 0.923076923076923,
        "mean": 0.923076923076923,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.7054014374088451,
        "sum_squared": 0.4975911878984648,
        "min": 0.7054014374088451,
        "max": 0.7054014374088451,
        "mean": 0.7054014374088451,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.5643211499270759,
        "sum_squared": 0.31845836025501734,
        "min": 0.5643211499270759,
        "max": 0.5643211499270759,
        "mean": 0.5643211499270759,
        "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": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 34,
        "sum_squared": 1156,
        "min": 34,
        "max": 34,
        "mean": 34.0,
        "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": 3589,
        "sum_squared": 12880921,
        "min": 3589,
        "max": 3589,
        "mean": 3589.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 0.27483654022216797,
        "sum_squared": 0.07553512384129135,
        "min": 0.27483654022216797,
        "max": 0.27483654022216797,
        "mean": 0.27483654022216797,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": "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": 3207,
        "sum_squared": 10284849,
        "min": 3207,
        "max": 3207,
        "mean": 3207.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.8442568778991699,
        "sum_squared": 0.7127696758800539,
        "min": 0.8442568778991699,
        "max": 0.8442568778991699,
        "mean": 0.8442568778991699,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": "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": 3605,
        "sum_squared": 12996025,
        "min": 3605,
        "max": 3605,
        "mean": 3605.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.8656706809997559,
        "sum_squared": 0.7493857279425811,
        "min": 0.8656706809997559,
        "max": 0.8656706809997559,
        "mean": 0.8656706809997559,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 0.5,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": "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": 3112,
        "sum_squared": 9684544,
        "min": 3112,
        "max": 3112,
        "mean": 3112.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.2510643005371094,
        "sum_squared": 0.06303328300418798,
        "min": 0.2510643005371094,
        "max": 0.2510643005371094,
        "mean": 0.2510643005371094,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "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": 0.1353352832366127,
        "sum_squared": 0.018315638888734182,
        "min": 0.1353352832366127,
        "max": 0.1353352832366127,
        "mean": 0.1353352832366127,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 3.01131000863463e-309,
        "sum_squared": 0.0,
        "min": 3.01131000863463e-309,
        "max": 3.01131000863463e-309,
        "mean": 3.01131000863463e-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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": "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": 3487,
        "sum_squared": 12159169,
        "min": 3487,
        "max": 3487,
        "mean": 3487.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.7961385250091553,
        "sum_squared": 0.6338365510037534,
        "min": 0.7961385250091553,
        "max": 0.7961385250091553,
        "mean": 0.7961385250091553,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "valid"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "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": 3276,
        "sum_squared": 10732176,
        "min": 3276,
        "max": 3276,
        "mean": 3276.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "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": 0.8257331848144531,
        "sum_squared": 0.6818352925038198,
        "min": 0.8257331848144531,
        "max": 0.8257331848144531,
        "mean": 0.8257331848144531,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5454545454545454,
        "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": 0.42857142857142855,
        "sum_squared": 0.18367346938775508,
        "min": 0.42857142857142855,
        "max": 0.42857142857142855,
        "mean": 0.42857142857142855,
        "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": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "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": 3174,
        "sum_squared": 10074276,
        "min": 3174,
        "max": 3174,
        "mean": 3174.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.7646636962890625,
        "sum_squared": 0.5847105684224516,
        "min": 0.7646636962890625,
        "max": 0.7646636962890625,
        "mean": 0.7646636962890625,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": "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": 3500,
        "sum_squared": 12250000,
        "min": 3500,
        "max": 3500,
        "mean": 3500.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 0.22931790351867676,
        "sum_squared": 0.05258670087420114,
        "min": 0.22931790351867676,
        "max": 0.22931790351867676,
        "mean": 0.22931790351867676,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 6,
        "sum_squared": 36,
        "min": 6,
        "max": 6,
        "mean": 6.0,
        "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": 3153,
        "sum_squared": 9941409,
        "min": 3153,
        "max": 3153,
        "mean": 3153.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.7623381614685059,
        "sum_squared": 0.5811594724311817,
        "min": 0.7623381614685059,
        "max": 0.7623381614685059,
        "mean": 0.7623381614685059,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5,
        "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": 0.049787068367863944,
        "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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": "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": 3585,
        "sum_squared": 12852225,
        "min": 3585,
        "max": 3585,
        "mean": 3585.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "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": 0.8195762634277344,
        "sum_squared": 0.671705251574167,
        "min": 0.8195762634277344,
        "max": 0.8195762634277344,
        "mean": 0.8195762634277344,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "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": "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": 3571,
        "sum_squared": 12752041,
        "min": 3571,
        "max": 3571,
        "mean": 3571.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.3450310230255127,
        "sum_squared": 0.11904640685003187,
        "min": 0.3450310230255127,
        "max": 0.3450310230255127,
        "mean": 0.3450310230255127,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": "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": 3053,
        "sum_squared": 9320809,
        "min": 3053,
        "max": 3053,
        "mean": 3053.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 6,
        "sum_squared": 36,
        "min": 6,
        "max": 6,
        "mean": 6.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 6,
        "sum_squared": 36,
        "min": 6,
        "max": 6,
        "mean": 6.0,
        "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": 0.8340392112731934,
        "sum_squared": 0.6956214059412105,
        "min": 0.8340392112731934,
        "max": 0.8340392112731934,
        "mean": 0.8340392112731934,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6666666666666666,
        "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": 0.75,
        "sum_squared": 0.5625,
        "min": 0.75,
        "max": 0.75,
        "mean": 0.75,
        "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": 6,
        "sum_squared": 36,
        "min": 6,
        "max": 6,
        "mean": 6.0,
        "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": "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": 3543,
        "sum_squared": 12552849,
        "min": 3543,
        "max": 3543,
        "mean": 3543.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.2490241527557373,
        "sum_squared": 0.06201302865571279,
        "min": 0.2490241527557373,
        "max": 0.2490241527557373,
        "mean": 0.2490241527557373,
        "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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": "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": 3442,
        "sum_squared": 11847364,
        "min": 3442,
        "max": 3442,
        "mean": 3442.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "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": 0.29587244987487793,
        "sum_squared": 0.08754050659496215,
        "min": 0.29587244987487793,
        "max": 0.29587244987487793,
        "mean": 0.29587244987487793,
        "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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 3,
        "sum_squared": 9,
        "min": 3,
        "max": 3,
        "mean": 3.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 14,
        "sum_squared": 196,
        "min": 14,
        "max": 14,
        "mean": 14.0,
        "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": 3871,
        "sum_squared": 14984641,
        "min": 3871,
        "max": 3871,
        "mean": 3871.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "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_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": 0.2909660339355469,
        "sum_squared": 0.08466123290418182,
        "min": 0.2909660339355469,
        "max": 0.2909660339355469,
        "mean": 0.2909660339355469,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "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": "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": 3132,
        "sum_squared": 9809424,
        "min": 3132,
        "max": 3132,
        "mean": 3132.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "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": 0.38845372200012207,
        "sum_squared": 0.15089629413574812,
        "min": 0.38845372200012207,
        "max": 0.38845372200012207,
        "mean": 0.38845372200012207,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2,
        "sum_squared": 0.04000000000000001,
        "min": 0.2,
        "max": 0.2,
        "mean": 0.2,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.15163266492815836,
        "sum_squared": 0.022992465073215146,
        "min": 0.15163266492815836,
        "max": 0.15163266492815836,
        "mean": 0.15163266492815836,
        "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": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 90,
        "sum_squared": 8100,
        "min": 90,
        "max": 90,
        "mean": 90.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  }
]