[
  {
    "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": 3739,
        "sum_squared": 13980121,
        "min": 3739,
        "max": 3739,
        "mean": 3739.0,
        "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": 1.5412235260009766,
        "sum_squared": 2.375369957098883,
        "min": 1.5412235260009766,
        "max": 1.5412235260009766,
        "mean": 1.5412235260009766,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5714285714285714,
        "sum_squared": 0.32653061224489793,
        "min": 0.5714285714285714,
        "max": 0.5714285714285714,
        "mean": 0.5714285714285714,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6,
        "sum_squared": 0.36,
        "min": 0.6,
        "max": 0.6,
        "mean": 0.6,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.25,
        "sum_squared": 0.0625,
        "min": 0.25,
        "max": 0.25,
        "mean": 0.25,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 34,
        "sum_squared": 1156,
        "min": 34,
        "max": 34,
        "mean": 34.0,
        "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": 3590,
        "sum_squared": 12888100,
        "min": 3590,
        "max": 3590,
        "mean": 3590.0,
        "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": 1.4449994564056396,
        "sum_squared": 2.088023429012594,
        "min": 1.4449994564056396,
        "max": 1.4449994564056396,
        "mean": 1.4449994564056396,
        "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": 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": "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": 3614,
        "sum_squared": 13060996,
        "min": 3614,
        "max": 3614,
        "mean": 3614.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.431703805923462,
        "sum_squared": 2.049775787895726,
        "min": 1.431703805923462,
        "max": 1.431703805923462,
        "mean": 1.431703805923462,
        "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": 0.36787944117144233,
        "sum_squared": 0.1353352832366127,
        "min": 0.36787944117144233,
        "max": 0.36787944117144233,
        "mean": 0.36787944117144233,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 8.18558927632837e-309,
        "sum_squared": 0.0,
        "min": 8.18558927632837e-309,
        "max": 8.18558927632837e-309,
        "mean": 8.18558927632837e-309,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "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": 3720,
        "sum_squared": 13838400,
        "min": 3720,
        "max": 3720,
        "mean": 3720.0,
        "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": 1.439913272857666,
        "sum_squared": 2.0733502333516753,
        "min": 1.439913272857666,
        "max": 1.439913272857666,
        "mean": 1.439913272857666,
        "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": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1340",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3233,
        "sum_squared": 10452289,
        "min": 3233,
        "max": 3233,
        "mean": 3233.0,
        "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": 1.559389591217041,
        "sum_squared": 2.4316958971960503,
        "min": 1.559389591217041,
        "max": 1.559389591217041,
        "mean": 1.559389591217041,
        "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": 0.7368421052631579,
        "sum_squared": 0.5429362880886426,
        "min": 0.7368421052631579,
        "max": 0.7368421052631579,
        "mean": 0.7368421052631579,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.64,
        "sum_squared": 0.4096,
        "min": 0.64,
        "max": 0.64,
        "mean": 0.64,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6923076923076923,
        "sum_squared": 0.4792899408284023,
        "min": 0.6923076923076923,
        "max": 0.6923076923076923,
        "mean": 0.6923076923076923,
        "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": 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": 65,
        "sum_squared": 4225,
        "min": 65,
        "max": 65,
        "mean": 65.0,
        "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": 3854,
        "sum_squared": 14853316,
        "min": 3854,
        "max": 3854,
        "mean": 3854.0,
        "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": 1.3893020153045654,
        "sum_squared": 1.930160089729327,
        "min": 1.3893020153045654,
        "max": 1.3893020153045654,
        "mean": 1.3893020153045654,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0,
        "sum_squared": 0.0,
        "min": 0.0,
        "max": 0.0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 18,
        "sum_squared": 324,
        "min": 18,
        "max": 18,
        "mean": 18.0,
        "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": 3528,
        "sum_squared": 12446784,
        "min": 3528,
        "max": 3528,
        "mean": 3528.0,
        "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": 1.3569517135620117,
        "sum_squared": 1.84131795293888,
        "min": 1.3569517135620117,
        "max": 1.3569517135620117,
        "mean": 1.3569517135620117,
        "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": 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": 17,
        "sum_squared": 289,
        "min": 17,
        "max": 17,
        "mean": 17.0,
        "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": 3889,
        "sum_squared": 15124321,
        "min": 3889,
        "max": 3889,
        "mean": 3889.0,
        "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": 1.297419548034668,
        "sum_squared": 1.683297483622482,
        "min": 1.297419548034668,
        "max": 1.297419548034668,
        "mean": 1.297419548034668,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0,
        "sum_squared": 0.0,
        "min": 0.0,
        "max": 0.0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1173",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3650,
        "sum_squared": 13322500,
        "min": 3650,
        "max": 3650,
        "mean": 3650.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.4805212020874023,
        "sum_squared": 2.191943029830327,
        "min": 1.4805212020874023,
        "max": 1.4805212020874023,
        "mean": 1.4805212020874023,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 28,
        "sum_squared": 784,
        "min": 28,
        "max": 28,
        "mean": 28.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1483",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3436,
        "sum_squared": 11806096,
        "min": 3436,
        "max": 3436,
        "mean": 3436.0,
        "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": 1.3946609497070312,
        "sum_squared": 1.9450791646377183,
        "min": 1.3946609497070312,
        "max": 1.3946609497070312,
        "mean": 1.3946609497070312,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0,
        "sum_squared": 0.0,
        "min": 0.0,
        "max": 0.0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 17,
        "sum_squared": 289,
        "min": 17,
        "max": 17,
        "mean": 17.0,
        "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": 3832,
        "sum_squared": 14684224,
        "min": 3832,
        "max": 3832,
        "mean": 3832.0,
        "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": 1.4525930881500244,
        "sum_squared": 2.1100266797412246,
        "min": 1.4525930881500244,
        "max": 1.4525930881500244,
        "mean": 1.4525930881500244,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 30,
        "sum_squared": 900,
        "min": 30,
        "max": 30,
        "mean": 30.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "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": 3949,
        "sum_squared": 15594601,
        "min": 3949,
        "max": 3949,
        "mean": 3949.0,
        "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": 1.3924665451049805,
        "sum_squared": 1.9389630792366006,
        "min": 1.3924665451049805,
        "max": 1.3924665451049805,
        "mean": 1.3924665451049805,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0,
        "sum_squared": 0.0,
        "min": 0.0,
        "max": 0.0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.10000000000000002,
        "sum_squared": 0.010000000000000004,
        "min": 0.10000000000000002,
        "max": 0.10000000000000002,
        "mean": 0.10000000000000002,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 47,
        "sum_squared": 2209,
        "min": 47,
        "max": 47,
        "mean": 47.0,
        "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": 3976,
        "sum_squared": 15808576,
        "min": 3976,
        "max": 3976,
        "mean": 3976.0,
        "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": 1.762603521347046,
        "sum_squared": 3.106771173465006,
        "min": 1.762603521347046,
        "max": 1.762603521347046,
        "mean": 1.762603521347046,
        "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": 0.8,
        "sum_squared": 0.6400000000000001,
        "min": 0.8,
        "max": 0.8,
        "mean": 0.8,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.7999999999999999,
        "sum_squared": 0.6399999999999999,
        "min": 0.7999999999999999,
        "max": 0.7999999999999999,
        "mean": 0.7999999999999999,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.7777777777777778,
        "sum_squared": 0.6049382716049383,
        "min": 0.7777777777777778,
        "max": 0.7777777777777778,
        "mean": 0.7777777777777778,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "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": "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": 39,
        "sum_squared": 1521,
        "min": 39,
        "max": 39,
        "mean": 39.0,
        "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": 3808,
        "sum_squared": 14500864,
        "min": 3808,
        "max": 3808,
        "mean": 3808.0,
        "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": 1.6217596530914307,
        "sum_squared": 2.6301043723952375,
        "min": 1.6217596530914307,
        "max": 1.6217596530914307,
        "mean": 1.6217596530914307,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0,
        "sum_squared": 0.0,
        "min": 0.0,
        "max": 0.0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 12,
        "sum_squared": 144,
        "min": 12,
        "max": 12,
        "mean": 12.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1529",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3519,
        "sum_squared": 12383361,
        "min": 3519,
        "max": 3519,
        "mean": 3519.0,
        "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": 0.7493264675140381,
        "sum_squared": 0.5614901549170668,
        "min": 0.7493264675140381,
        "max": 0.7493264675140381,
        "mean": 0.7493264675140381,
        "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": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1209",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3737,
        "sum_squared": 13965169,
        "min": 3737,
        "max": 3737,
        "mean": 3737.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.5442373752593994,
        "sum_squared": 2.384669071148039,
        "min": 1.5442373752593994,
        "max": 1.5442373752593994,
        "mean": 1.5442373752593994,
        "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": 0.6065306597126334,
        "sum_squared": 0.36787944117144233,
        "min": 0.6065306597126334,
        "max": 0.6065306597126334,
        "mean": 0.6065306597126334,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "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": 3479,
        "sum_squared": 12103441,
        "min": 3479,
        "max": 3479,
        "mean": 3479.0,
        "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": 0.7757816314697266,
        "sum_squared": 0.6018371397258306,
        "min": 0.7757816314697266,
        "max": 0.7757816314697266,
        "mean": 0.7757816314697266,
        "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": 0.4,
        "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": 0.27590958087858175,
        "sum_squared": 0.07612609682059464,
        "min": 0.27590958087858175,
        "max": 0.27590958087858175,
        "mean": 0.27590958087858175,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 8.18558927632837e-309,
        "sum_squared": 0.0,
        "min": 8.18558927632837e-309,
        "max": 8.18558927632837e-309,
        "mean": 8.18558927632837e-309,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 21,
        "sum_squared": 441,
        "min": 21,
        "max": 21,
        "mean": 21.0,
        "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": 3372,
        "sum_squared": 11370384,
        "min": 3372,
        "max": 3372,
        "mean": 3372.0,
        "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": 1.752368688583374,
        "sum_squared": 3.070796020727414,
        "min": 1.752368688583374,
        "max": 1.752368688583374,
        "mean": 1.752368688583374,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 20,
        "sum_squared": 400,
        "min": 20,
        "max": 20,
        "mean": 20.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "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": 3448,
        "sum_squared": 11888704,
        "min": 3448,
        "max": 3448,
        "mean": 3448.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 1.5551815032958984,
        "sum_squared": 2.4185895081936906,
        "min": 1.5551815032958984,
        "max": 1.5551815032958984,
        "mean": 1.5551815032958984,
        "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": 0.47058823529411764,
        "sum_squared": 0.22145328719723184,
        "min": 0.47058823529411764,
        "max": 0.47058823529411764,
        "mean": 0.47058823529411764,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.375,
        "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": 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": 70,
        "sum_squared": 4900,
        "min": 70,
        "max": 70,
        "mean": 70.0,
        "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": 3874,
        "sum_squared": 15007876,
        "min": 3874,
        "max": 3874,
        "mean": 3874.0,
        "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": 1.3574583530426025,
        "sum_squared": 1.842693180245135,
        "min": 1.3574583530426025,
        "max": 1.3574583530426025,
        "mean": 1.3574583530426025,
        "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": "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": 3658,
        "sum_squared": 13380964,
        "min": 3658,
        "max": 3658,
        "mean": 3658.0,
        "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": 1.3098211288452148,
        "sum_squared": 1.715631389569353,
        "min": 1.3098211288452148,
        "max": 1.3098211288452148,
        "mean": 1.3098211288452148,
        "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": 30,
        "sum_squared": 900,
        "min": 30,
        "max": 30,
        "mean": 30.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1335",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3898,
        "sum_squared": 15194404,
        "min": 3898,
        "max": 3898,
        "mean": 3898.0,
        "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": 1.378843069076538,
        "sum_squared": 1.9012082091404068,
        "min": 1.378843069076538,
        "max": 1.378843069076538,
        "mean": 1.378843069076538,
        "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": 9,
        "sum_squared": 81,
        "min": 9,
        "max": 9,
        "mean": 9.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1337",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3638,
        "sum_squared": 13235044,
        "min": 3638,
        "max": 3638,
        "mean": 3638.0,
        "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": 1.3743243217468262,
        "sum_squared": 1.8887673413448738,
        "min": 1.3743243217468262,
        "max": 1.3743243217468262,
        "mean": 1.3743243217468262,
        "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": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1114",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3370,
        "sum_squared": 11356900,
        "min": 3370,
        "max": 3370,
        "mean": 3370.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.2421355247497559,
        "sum_squared": 1.5429006618453514,
        "min": 1.2421355247497559,
        "max": 1.2421355247497559,
        "mean": 1.2421355247497559,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1198",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3780,
        "sum_squared": 14288400,
        "min": 3780,
        "max": 3780,
        "mean": 3780.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.278876543045044,
        "sum_squared": 1.6355252123508421,
        "min": 1.278876543045044,
        "max": 1.278876543045044,
        "mean": 1.278876543045044,
        "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": 0.8421052631578947,
        "sum_squared": 0.709141274238227,
        "min": 0.8421052631578947,
        "max": 0.8421052631578947,
        "mean": 0.8421052631578947,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.7368421052631579,
        "sum_squared": 0.5429362880886426,
        "min": 0.7368421052631579,
        "max": 0.7368421052631579,
        "mean": 0.7368421052631579,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.7165313105737893,
        "sum_squared": 0.513417119032592,
        "min": 0.7165313105737893,
        "max": 0.7165313105737893,
        "mean": 0.7165313105737893,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 0.35826565528689464,
        "sum_squared": 0.128354279758148,
        "min": 0.35826565528689464,
        "max": 0.35826565528689464,
        "mean": 0.35826565528689464,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 52,
        "sum_squared": 2704,
        "min": 52,
        "max": 52,
        "mean": 52.0,
        "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": 3325,
        "sum_squared": 11055625,
        "min": 3325,
        "max": 3325,
        "mean": 3325.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.3955211639404297,
        "sum_squared": 1.9474793190056516,
        "min": 1.3955211639404297,
        "max": 1.3955211639404297,
        "mean": 1.3955211639404297,
        "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": 0.75,
        "sum_squared": 0.5625,
        "min": 0.75,
        "max": 0.75,
        "mean": 0.75,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 20,
        "sum_squared": 400,
        "min": 20,
        "max": 20,
        "mean": 20.0,
        "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": 3720,
        "sum_squared": 13838400,
        "min": 3720,
        "max": 3720,
        "mean": 3720.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.3848824501037598,
        "sum_squared": 1.9178994006053927,
        "min": 1.3848824501037598,
        "max": 1.3848824501037598,
        "mean": 1.3848824501037598,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1273",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3466,
        "sum_squared": 12013156,
        "min": 3466,
        "max": 3466,
        "mean": 3466.0,
        "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": 1.4690089225769043,
        "sum_squared": 2.157987214610557,
        "min": 1.4690089225769043,
        "max": 1.4690089225769043,
        "mean": 1.4690089225769043,
        "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": 0.8461538461538461,
        "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": 0.7333333333333333,
        "sum_squared": 0.5377777777777777,
        "min": 0.7333333333333333,
        "max": 0.7333333333333333,
        "mean": 0.7333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "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": "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": 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": 3509,
        "sum_squared": 12313081,
        "min": 3509,
        "max": 3509,
        "mean": 3509.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.2607371807098389,
        "sum_squared": 1.589458238824193,
        "min": 1.2607371807098389,
        "max": 1.2607371807098389,
        "mean": 1.2607371807098389,
        "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": 4,
        "sum_squared": 16,
        "min": 4,
        "max": 4,
        "mean": 4.0,
        "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": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "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": 3989,
        "sum_squared": 15912121,
        "min": 3989,
        "max": 3989,
        "mean": 3989.0,
        "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": 1.610536813735962,
        "sum_squared": 2.5938288283987845,
        "min": 1.610536813735962,
        "max": 1.610536813735962,
        "mean": 1.610536813735962,
        "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": 0.5333333333333333,
        "sum_squared": 0.28444444444444444,
        "min": 0.5333333333333333,
        "max": 0.5333333333333333,
        "mean": 0.5333333333333333,
        "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": 0.3333333333333333,
        "sum_squared": 0.1111111111111111,
        "min": 0.3333333333333333,
        "max": 0.3333333333333333,
        "mean": 0.3333333333333333,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 50,
        "sum_squared": 2500,
        "min": 50,
        "max": 50,
        "mean": 50.0,
        "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": 3863,
        "sum_squared": 14922769,
        "min": 3863,
        "max": 3863,
        "mean": 3863.0,
        "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": 1.5502355098724365,
        "sum_squared": 2.4032301360694532,
        "min": 1.5502355098724365,
        "max": 1.5502355098724365,
        "mean": 1.5502355098724365,
        "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": 0.8888888888888888,
        "sum_squared": 0.7901234567901234,
        "min": 0.8888888888888888,
        "max": 0.8888888888888888,
        "mean": 0.8888888888888888,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.8333333333333334,
        "sum_squared": 0.6944444444444445,
        "min": 0.8333333333333334,
        "max": 0.8333333333333334,
        "mean": 0.8333333333333334,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "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": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 1.8834843577582417e-308,
        "sum_squared": 0.0,
        "min": 1.8834843577582417e-308,
        "max": 1.8834843577582417e-308,
        "mean": 1.8834843577582417e-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": 33,
        "sum_squared": 1089,
        "min": 33,
        "max": 33,
        "mean": 33.0,
        "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": 3816,
        "sum_squared": 14561856,
        "min": 3816,
        "max": 3816,
        "mean": 3816.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.4730842113494873,
        "sum_squared": 2.169977093727141,
        "min": 1.4730842113494873,
        "max": 1.4730842113494873,
        "mean": 1.4730842113494873,
        "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": 0.36787944117144233,
        "sum_squared": 0.1353352832366127,
        "min": 0.36787944117144233,
        "max": 0.36787944117144233,
        "mean": 0.36787944117144233,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "valid"
        },
        "count": 1,
        "sum": 8.18558927632837e-309,
        "sum_squared": 0.0,
        "min": 8.18558927632837e-309,
        "max": 8.18558927632837e-309,
        "mean": 8.18558927632837e-309,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "valid"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 10,
        "sum_squared": 100,
        "min": 10,
        "max": 10,
        "mean": 10.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1463",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3902,
        "sum_squared": 15225604,
        "min": 3902,
        "max": 3902,
        "mean": 3902.0,
        "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": 1.3766818046569824,
        "sum_squared": 1.895252791273606,
        "min": 1.3766818046569824,
        "max": 1.3766818046569824,
        "mean": 1.3766818046569824,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 15,
        "sum_squared": 225,
        "min": 15,
        "max": 15,
        "mean": 15.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1112",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "valid"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 3800,
        "sum_squared": 14440000,
        "min": 3800,
        "max": 3800,
        "mean": 3800.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.5397047996520996,
        "sum_squared": 2.370690870071712,
        "min": 1.5397047996520996,
        "max": 1.5397047996520996,
        "mean": 1.5397047996520996,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 35,
        "sum_squared": 1225,
        "min": 35,
        "max": 35,
        "mean": 35.0,
        "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": 3793,
        "sum_squared": 14386849,
        "min": 3793,
        "max": 3793,
        "mean": 3793.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 1.445272445678711,
        "sum_squared": 2.0888124422381225,
        "min": 1.445272445678711,
        "max": 1.445272445678711,
        "mean": 1.445272445678711,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1256",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3418,
        "sum_squared": 11682724,
        "min": 3418,
        "max": 3418,
        "mean": 3418.0,
        "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": 1.3753232955932617,
        "sum_squared": 1.8915141674015103,
        "min": 1.3753232955932617,
        "max": 1.3753232955932617,
        "mean": 1.3753232955932617,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1499",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3819,
        "sum_squared": 14584761,
        "min": 3819,
        "max": 3819,
        "mean": 3819.0,
        "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": 1.3282570838928223,
        "sum_squared": 1.7642668809114639,
        "min": 1.3282570838928223,
        "max": 1.3282570838928223,
        "mean": 1.3282570838928223,
        "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": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "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": 3314,
        "sum_squared": 10982596,
        "min": 3314,
        "max": 3314,
        "mean": 3314.0,
        "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": 1.3524811267852783,
        "sum_squared": 1.829205198310376,
        "min": 1.3524811267852783,
        "max": 1.3524811267852783,
        "mean": 1.3524811267852783,
        "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": 0.8,
        "sum_squared": 0.6400000000000001,
        "min": 0.8,
        "max": 0.8,
        "mean": 0.8,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.6065306597126334,
        "sum_squared": 0.36787944117144233,
        "min": 0.6065306597126334,
        "max": 0.6065306597126334,
        "mean": 0.6065306597126334,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "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": "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": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "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": 3699,
        "sum_squared": 13682601,
        "min": 3699,
        "max": 3699,
        "mean": 3699.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_output_tokens",
          "split": "valid"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "training_co2_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "training_energy_cost",
          "split": "valid"
        },
        "count": 0,
        "sum": 0,
        "sum_squared": 0
      },
      {
        "name": {
          "name": "inference_runtime",
          "split": "valid"
        },
        "count": 1,
        "sum": 1.3125126361846924,
        "sum_squared": 1.7226894201444907,
        "min": 1.3125126361846924,
        "max": 1.3125126361846924,
        "mean": 1.3125126361846924,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "valid"
        },
        "count": 1,
        "sum": 17,
        "sum_squared": 289,
        "min": 17,
        "max": 17,
        "mean": 17.0,
        "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": 3489,
        "sum_squared": 12173121,
        "min": 3489,
        "max": 3489,
        "mean": 3489.0,
        "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": 1.355616807937622,
        "sum_squared": 1.8376969299629877,
        "min": 1.355616807937622,
        "max": 1.355616807937622,
        "mean": 1.355616807937622,
        "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": 0.5,
        "sum_squared": 0.25,
        "min": 0.5,
        "max": 0.5,
        "mean": 0.5,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.2250738585072626e-308,
        "sum_squared": 0.0,
        "min": 2.2250738585072626e-308,
        "max": 2.2250738585072626e-308,
        "mean": 2.2250738585072626e-308,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 3406,
        "sum_squared": 11600836,
        "min": 3406,
        "max": 3406,
        "mean": 3406.0,
        "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": 1.2535169124603271,
        "sum_squared": 1.5713046498240715,
        "min": 1.2535169124603271,
        "max": 1.2535169124603271,
        "mean": 1.2535169124603271,
        "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": 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": 11,
        "sum_squared": 121,
        "min": 11,
        "max": 11,
        "mean": 11.0,
        "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": 3725,
        "sum_squared": 13875625,
        "min": 3725,
        "max": 3725,
        "mean": 3725.0,
        "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": 1.3806896209716797,
        "sum_squared": 1.9063038294589205,
        "min": 1.3806896209716797,
        "max": 1.3806896209716797,
        "mean": 1.3806896209716797,
        "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": "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": 3374,
        "sum_squared": 11383876,
        "min": 3374,
        "max": 3374,
        "mean": 3374.0,
        "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": 1.4275085926055908,
        "sum_squared": 2.0377807819627947,
        "min": 1.4275085926055908,
        "max": 1.4275085926055908,
        "mean": 1.4275085926055908,
        "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": 0.75,
        "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": 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": 19,
        "sum_squared": 361,
        "min": 19,
        "max": 19,
        "mean": 19.0,
        "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": 3883,
        "sum_squared": 15077689,
        "min": 3883,
        "max": 3883,
        "mean": 3883.0,
        "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": 1.3695757389068604,
        "sum_squared": 1.8757377046022725,
        "min": 1.3695757389068604,
        "max": 1.3695757389068604,
        "mean": 1.3695757389068604,
        "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": 16,
        "sum_squared": 256,
        "min": 16,
        "max": 16,
        "mean": 16.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1495",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3781,
        "sum_squared": 14295961,
        "min": 3781,
        "max": 3781,
        "mean": 3781.0,
        "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": 1.3324410915374756,
        "sum_squared": 1.7753992624175794,
        "min": 1.3324410915374756,
        "max": 1.3324410915374756,
        "mean": 1.3324410915374756,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.2857142857142857,
        "sum_squared": 0.08163265306122448,
        "min": 0.2857142857142857,
        "max": 0.2857142857142857,
        "mean": 0.2857142857142857,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0009118819655545162,
        "sum_squared": 8.31528719103568e-07,
        "min": 0.0009118819655545162,
        "max": 0.0009118819655545162,
        "mean": 0.0009118819655545162,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 2.0290047235994e-311,
        "sum_squared": 0.0,
        "min": 2.0290047235994e-311,
        "max": 2.0290047235994e-311,
        "mean": 2.0290047235994e-311,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1445",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3260,
        "sum_squared": 10627600,
        "min": 3260,
        "max": 3260,
        "mean": 3260.0,
        "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": 0.8475954532623291,
        "sum_squared": 0.7184180523909731,
        "min": 0.8475954532623291,
        "max": 0.8475954532623291,
        "mean": 0.8475954532623291,
        "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": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1507",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3786,
        "sum_squared": 14333796,
        "min": 3786,
        "max": 3786,
        "mean": 3786.0,
        "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": 1.6479179859161377,
        "sum_squared": 2.7156336883059,
        "min": 1.6479179859161377,
        "max": 1.6479179859161377,
        "mean": 1.6479179859161377,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 8,
        "sum_squared": 64,
        "min": 8,
        "max": 8,
        "mean": 8.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1414",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3660,
        "sum_squared": 13395600,
        "min": 3660,
        "max": 3660,
        "mean": 3660.0,
        "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": 1.601102590560913,
        "sum_squared": 2.563529505500867,
        "min": 1.601102590560913,
        "max": 1.601102590560913,
        "mean": 1.601102590560913,
        "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,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.0,
        "sum_squared": 0.0,
        "min": 0.0,
        "max": 0.0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 13,
        "sum_squared": 169,
        "min": 13,
        "max": 13,
        "mean": 13.0,
        "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": 3570,
        "sum_squared": 12744900,
        "min": 3570,
        "max": 3570,
        "mean": 3570.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_completion_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 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": 1.5870811939239502,
        "sum_squared": 2.518826716107071,
        "min": 1.5870811939239502,
        "max": 1.5870811939239502,
        "mean": 1.5870811939239502,
        "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,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_1",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "bleu_4",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "logprob",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_perplexity_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 7,
        "sum_squared": 49,
        "min": 7,
        "max": 7,
        "mean": 7.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  },
  {
    "instance_id": "id1555",
    "train_trial_index": 0,
    "stats": [
      {
        "name": {
          "name": "num_references",
          "split": "test"
        },
        "count": 1,
        "sum": 2,
        "sum_squared": 4,
        "min": 2,
        "max": 2,
        "mean": 2.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_trials",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_prompt_tokens",
          "split": "test"
        },
        "count": 1,
        "sum": 3324,
        "sum_squared": 11048976,
        "min": 3324,
        "max": 3324,
        "mean": 3324.0,
        "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": 0.8100006580352783,
        "sum_squared": 0.6561010660175839,
        "min": 0.8100006580352783,
        "max": 0.8100006580352783,
        "mean": 0.8100006580352783,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "batch_size",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_length",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_stop",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_endoftext",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "finish_reason_unknown",
          "split": "test"
        },
        "count": 1,
        "sum": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_train_instances",
          "split": "test"
        },
        "count": 1,
        "sum": 5,
        "sum_squared": 25,
        "min": 5,
        "max": 5,
        "mean": 5.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "prompt_truncated",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "max_prob",
          "split": "test"
        },
        "count": 1,
        "sum": 1.0,
        "sum_squared": 1.0,
        "min": 1.0,
        "max": 1.0,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "quasi_exact_match",
          "split": "test"
        },
        "count": 1,
        "sum": 0,
        "sum_squared": 0,
        "min": 0,
        "max": 0,
        "mean": 0.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "f1_score",
          "split": "test"
        },
        "count": 1,
        "sum": 0.3076923076923077,
        "sum_squared": 0.09467455621301776,
        "min": 0.3076923076923077,
        "max": 0.3076923076923077,
        "mean": 0.3076923076923077,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "rouge_l",
          "split": "test"
        },
        "count": 1,
        "sum": 0.26666666666666666,
        "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": 0.0909090909090909,
        "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": 1,
        "sum_squared": 1,
        "min": 1,
        "max": 1,
        "mean": 1.0,
        "variance": 0.0,
        "stddev": 0.0
      },
      {
        "name": {
          "name": "num_bytes",
          "split": "test"
        },
        "count": 1,
        "sum": 60,
        "sum_squared": 3600,
        "min": 60,
        "max": 60,
        "mean": 60.0,
        "variance": 0.0,
        "stddev": 0.0
      }
    ]
  }
]