[
  {
    "name": {
      "name": "num_references",
      "split": "test"
    },
    "count": 1,
    "sum": 2.0,
    "sum_squared": 4.0,
    "min": 2.0,
    "max": 2.0,
    "mean": 2.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_trials",
      "split": "test"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_prompt_tokens",
      "split": "test"
    },
    "count": 1,
    "sum": 3464.5641025641025,
    "sum_squared": 12003204.420775805,
    "min": 3464.5641025641025,
    "max": 3464.5641025641025,
    "mean": 3464.5641025641025,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "test"
    },
    "count": 1,
    "sum": 6.256410256410256,
    "sum_squared": 39.14266929651545,
    "min": 6.256410256410256,
    "max": 6.256410256410256,
    "mean": 6.256410256410256,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "test"
    },
    "count": 1,
    "sum": 6.256410256410256,
    "sum_squared": 39.14266929651545,
    "min": 6.256410256410256,
    "max": 6.256410256410256,
    "mean": 6.256410256410256,
    "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": 0.6167610608614408,
    "sum_squared": 0.38039420619492986,
    "min": 0.6167610608614408,
    "max": 0.6167610608614408,
    "mean": 0.6167610608614408,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "batch_size",
      "split": "test"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "finish_reason_length",
      "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": "finish_reason_stop",
      "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": "finish_reason_endoftext",
      "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": "finish_reason_unknown",
      "split": "test"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_instances",
      "split": "test"
    },
    "count": 1,
    "sum": 4.948717948717949,
    "sum_squared": 24.489809335963184,
    "min": 4.948717948717949,
    "max": 4.948717948717949,
    "mean": 4.948717948717949,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prompt_truncated",
      "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": "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": 0.23076923076923078,
    "sum_squared": 0.053254437869822494,
    "min": 0.23076923076923078,
    "max": 0.23076923076923078,
    "mean": 0.23076923076923078,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "test"
    },
    "count": 1,
    "sum": 0.4358974358974359,
    "sum_squared": 0.19000657462195925,
    "min": 0.4358974358974359,
    "max": 0.4358974358974359,
    "mean": 0.4358974358974359,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "test"
    },
    "count": 1,
    "sum": 0.7174581326165038,
    "sum_squared": 0.5147461720575608,
    "min": 0.7174581326165038,
    "max": 0.7174581326165038,
    "mean": 0.7174581326165038,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "test"
    },
    "count": 1,
    "sum": 0.7193741462504141,
    "sum_squared": 0.5174991622935121,
    "min": 0.7193741462504141,
    "max": 0.7193741462504141,
    "mean": 0.7193741462504141,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "test"
    },
    "count": 1,
    "sum": 0.5221346013461805,
    "sum_squared": 0.27262454192293484,
    "min": 0.5221346013461805,
    "max": 0.5221346013461805,
    "mean": 0.5221346013461805,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "test"
    },
    "count": 1,
    "sum": 0.06503944773175543,
    "sum_squared": 0.004230129761251747,
    "min": 0.06503944773175543,
    "max": 0.06503944773175543,
    "mean": 0.06503944773175543,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "logprob",
      "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": "num_perplexity_tokens",
      "split": "test"
    },
    "count": 1,
    "sum": 6.256410256410256,
    "sum_squared": 39.14266929651545,
    "min": 6.256410256410256,
    "max": 6.256410256410256,
    "mean": 6.256410256410256,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "test"
    },
    "count": 1,
    "sum": 27.58974358974359,
    "sum_squared": 761.1939513477976,
    "min": 27.58974358974359,
    "max": 27.58974358974359,
    "mean": 27.58974358974359,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_references",
      "split": "valid"
    },
    "count": 1,
    "sum": 2.0,
    "sum_squared": 4.0,
    "min": 2.0,
    "max": 2.0,
    "mean": 2.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_trials",
      "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": "num_prompt_tokens",
      "split": "valid"
    },
    "count": 1,
    "sum": 3464.090909090909,
    "sum_squared": 11999925.82644628,
    "min": 3464.090909090909,
    "max": 3464.090909090909,
    "mean": 3464.090909090909,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "valid"
    },
    "count": 1,
    "sum": 5.2727272727272725,
    "sum_squared": 27.801652892561982,
    "min": 5.2727272727272725,
    "max": 5.2727272727272725,
    "mean": 5.2727272727272725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "valid"
    },
    "count": 1,
    "sum": 5.2727272727272725,
    "sum_squared": 27.801652892561982,
    "min": 5.2727272727272725,
    "max": 5.2727272727272725,
    "mean": 5.2727272727272725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "training_co2_cost",
      "split": "valid"
    },
    "count": 0,
    "sum": 0,
    "sum_squared": 0
  },
  {
    "name": {
      "name": "training_energy_cost",
      "split": "valid"
    },
    "count": 0,
    "sum": 0,
    "sum_squared": 0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.7053152864629572,
    "sum_squared": 0.4974696533183233,
    "min": 0.7053152864629572,
    "max": 0.7053152864629572,
    "mean": 0.7053152864629572,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "batch_size",
      "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": "finish_reason_length",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "finish_reason_stop",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "finish_reason_endoftext",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "finish_reason_unknown",
      "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": "num_train_instances",
      "split": "valid"
    },
    "count": 1,
    "sum": 5.0,
    "sum_squared": 25.0,
    "min": 5.0,
    "max": 5.0,
    "mean": 5.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prompt_truncated",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "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": 0.18181818181818182,
    "sum_squared": 0.03305785123966942,
    "min": 0.18181818181818182,
    "max": 0.18181818181818182,
    "mean": 0.18181818181818182,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.36363636363636365,
    "sum_squared": 0.1322314049586777,
    "min": 0.36363636363636365,
    "max": 0.36363636363636365,
    "mean": 0.36363636363636365,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.6914600550964187,
    "sum_squared": 0.4781170077939424,
    "min": 0.6914600550964187,
    "max": 0.6914600550964187,
    "mean": 0.6914600550964187,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.7630924630924631,
    "sum_squared": 0.5823101072285223,
    "min": 0.7630924630924631,
    "max": 0.7630924630924631,
    "mean": 0.7630924630924631,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.47949041914142077,
    "sum_squared": 0.22991106204841535,
    "min": 0.47949041914142077,
    "max": 0.47949041914142077,
    "mean": 0.47949041914142077,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.13882378668947293,
    "sum_squared": 0.019272043750804282,
    "min": 0.13882378668947293,
    "max": 0.13882378668947293,
    "mean": 0.13882378668947293,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "logprob",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_perplexity_tokens",
      "split": "valid"
    },
    "count": 1,
    "sum": 5.2727272727272725,
    "sum_squared": 27.801652892561982,
    "min": 5.2727272727272725,
    "max": 5.2727272727272725,
    "mean": 5.2727272727272725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "valid"
    },
    "count": 1,
    "sum": 20.454545454545453,
    "sum_squared": 418.3884297520661,
    "min": 20.454545454545453,
    "max": 20.454545454545453,
    "mean": 20.454545454545453,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "perplexity",
      "split": "test"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bits_per_byte",
      "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": "logprob_per_byte",
      "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": "perplexity",
      "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": "bits_per_byte",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "logprob_per_byte",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_references",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 2.0,
    "sum_squared": 4.0,
    "min": 2.0,
    "max": 2.0,
    "mean": 2.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_references",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 2.0,
    "sum_squared": 4.0,
    "min": 2.0,
    "max": 2.0,
    "mean": 2.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_trials",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_trials",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_prompt_tokens",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 3464.5641025641025,
    "sum_squared": 12003204.420775805,
    "min": 3464.5641025641025,
    "max": 3464.5641025641025,
    "mean": 3464.5641025641025,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_prompt_tokens",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 3464.5641025641025,
    "sum_squared": 12003204.420775805,
    "min": 3464.5641025641025,
    "max": 3464.5641025641025,
    "mean": 3464.5641025641025,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 6.256410256410256,
    "sum_squared": 39.14266929651545,
    "min": 6.256410256410256,
    "max": 6.256410256410256,
    "mean": 6.256410256410256,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 6.256410256410256,
    "sum_squared": 39.14266929651545,
    "min": 6.256410256410256,
    "max": 6.256410256410256,
    "mean": 6.256410256410256,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 6.256410256410256,
    "sum_squared": 39.14266929651545,
    "min": 6.256410256410256,
    "max": 6.256410256410256,
    "mean": 6.256410256410256,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 6.256410256410256,
    "sum_squared": 39.14266929651545,
    "min": 6.256410256410256,
    "max": 6.256410256410256,
    "mean": 6.256410256410256,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.6167610608614408,
    "sum_squared": 0.38039420619492986,
    "min": 0.6167610608614408,
    "max": 0.6167610608614408,
    "mean": 0.6167610608614408,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.6167610608614408,
    "sum_squared": 0.38039420619492986,
    "min": 0.6167610608614408,
    "max": 0.6167610608614408,
    "mean": 0.6167610608614408,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "batch_size",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "batch_size",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "finish_reason_length",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_length",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_stop",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_stop",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_endoftext",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_endoftext",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_unknown",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "finish_reason_unknown",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_instances",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 4.948717948717949,
    "sum_squared": 24.489809335963184,
    "min": 4.948717948717949,
    "max": 4.948717948717949,
    "mean": 4.948717948717949,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_instances",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 4.948717948717949,
    "sum_squared": 24.489809335963184,
    "min": 4.948717948717949,
    "max": 4.948717948717949,
    "mean": 4.948717948717949,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prompt_truncated",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "prompt_truncated",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "max_prob",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "max_prob",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "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",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.23076923076923078,
    "sum_squared": 0.053254437869822494,
    "min": 0.23076923076923078,
    "max": 0.23076923076923078,
    "mean": 0.23076923076923078,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.23076923076923078,
    "sum_squared": 0.053254437869822494,
    "min": 0.23076923076923078,
    "max": 0.23076923076923078,
    "mean": 0.23076923076923078,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.4358974358974359,
    "sum_squared": 0.19000657462195925,
    "min": 0.4358974358974359,
    "max": 0.4358974358974359,
    "mean": 0.4358974358974359,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.4358974358974359,
    "sum_squared": 0.19000657462195925,
    "min": 0.4358974358974359,
    "max": 0.4358974358974359,
    "mean": 0.4358974358974359,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7174581326165038,
    "sum_squared": 0.5147461720575608,
    "min": 0.7174581326165038,
    "max": 0.7174581326165038,
    "mean": 0.7174581326165038,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7174581326165038,
    "sum_squared": 0.5147461720575608,
    "min": 0.7174581326165038,
    "max": 0.7174581326165038,
    "mean": 0.7174581326165038,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7193741462504141,
    "sum_squared": 0.5174991622935121,
    "min": 0.7193741462504141,
    "max": 0.7193741462504141,
    "mean": 0.7193741462504141,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7193741462504141,
    "sum_squared": 0.5174991622935121,
    "min": 0.7193741462504141,
    "max": 0.7193741462504141,
    "mean": 0.7193741462504141,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.5221346013461805,
    "sum_squared": 0.27262454192293484,
    "min": 0.5221346013461805,
    "max": 0.5221346013461805,
    "mean": 0.5221346013461805,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.5221346013461805,
    "sum_squared": 0.27262454192293484,
    "min": 0.5221346013461805,
    "max": 0.5221346013461805,
    "mean": 0.5221346013461805,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.06503944773175543,
    "sum_squared": 0.004230129761251747,
    "min": 0.06503944773175543,
    "max": 0.06503944773175543,
    "mean": 0.06503944773175543,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.06503944773175543,
    "sum_squared": 0.004230129761251747,
    "min": 0.06503944773175543,
    "max": 0.06503944773175543,
    "mean": 0.06503944773175543,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "logprob",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "logprob",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "num_perplexity_tokens",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 6.256410256410256,
    "sum_squared": 39.14266929651545,
    "min": 6.256410256410256,
    "max": 6.256410256410256,
    "mean": 6.256410256410256,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_perplexity_tokens",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 6.256410256410256,
    "sum_squared": 39.14266929651545,
    "min": 6.256410256410256,
    "max": 6.256410256410256,
    "mean": 6.256410256410256,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 27.58974358974359,
    "sum_squared": 761.1939513477976,
    "min": 27.58974358974359,
    "max": 27.58974358974359,
    "mean": 27.58974358974359,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 27.58974358974359,
    "sum_squared": 761.1939513477976,
    "min": 27.58974358974359,
    "max": 27.58974358974359,
    "mean": 27.58974358974359,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_references",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 2.0,
    "sum_squared": 4.0,
    "min": 2.0,
    "max": 2.0,
    "mean": 2.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_references",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 2.0,
    "sum_squared": 4.0,
    "min": 2.0,
    "max": 2.0,
    "mean": 2.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_trials",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_trials",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_prompt_tokens",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 3464.090909090909,
    "sum_squared": 11999925.82644628,
    "min": 3464.090909090909,
    "max": 3464.090909090909,
    "mean": 3464.090909090909,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_prompt_tokens",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 3464.090909090909,
    "sum_squared": 11999925.82644628,
    "min": 3464.090909090909,
    "max": 3464.090909090909,
    "mean": 3464.090909090909,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 5.2727272727272725,
    "sum_squared": 27.801652892561982,
    "min": 5.2727272727272725,
    "max": 5.2727272727272725,
    "mean": 5.2727272727272725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 5.2727272727272725,
    "sum_squared": 27.801652892561982,
    "min": 5.2727272727272725,
    "max": 5.2727272727272725,
    "mean": 5.2727272727272725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 5.2727272727272725,
    "sum_squared": 27.801652892561982,
    "min": 5.2727272727272725,
    "max": 5.2727272727272725,
    "mean": 5.2727272727272725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 5.2727272727272725,
    "sum_squared": 27.801652892561982,
    "min": 5.2727272727272725,
    "max": 5.2727272727272725,
    "mean": 5.2727272727272725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7053152864629572,
    "sum_squared": 0.4974696533183233,
    "min": 0.7053152864629572,
    "max": 0.7053152864629572,
    "mean": 0.7053152864629572,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7053152864629572,
    "sum_squared": 0.4974696533183233,
    "min": 0.7053152864629572,
    "max": 0.7053152864629572,
    "mean": 0.7053152864629572,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "batch_size",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "batch_size",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "finish_reason_length",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_length",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_stop",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_stop",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_endoftext",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_endoftext",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "finish_reason_unknown",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "finish_reason_unknown",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_instances",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 5.0,
    "sum_squared": 25.0,
    "min": 5.0,
    "max": 5.0,
    "mean": 5.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_train_instances",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 5.0,
    "sum_squared": 25.0,
    "min": 5.0,
    "max": 5.0,
    "mean": 5.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prompt_truncated",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "prompt_truncated",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "max_prob",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "max_prob",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "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",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.18181818181818182,
    "sum_squared": 0.03305785123966942,
    "min": 0.18181818181818182,
    "max": 0.18181818181818182,
    "mean": 0.18181818181818182,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.18181818181818182,
    "sum_squared": 0.03305785123966942,
    "min": 0.18181818181818182,
    "max": 0.18181818181818182,
    "mean": 0.18181818181818182,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.36363636363636365,
    "sum_squared": 0.1322314049586777,
    "min": 0.36363636363636365,
    "max": 0.36363636363636365,
    "mean": 0.36363636363636365,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.36363636363636365,
    "sum_squared": 0.1322314049586777,
    "min": 0.36363636363636365,
    "max": 0.36363636363636365,
    "mean": 0.36363636363636365,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.6914600550964187,
    "sum_squared": 0.4781170077939424,
    "min": 0.6914600550964187,
    "max": 0.6914600550964187,
    "mean": 0.6914600550964187,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.6914600550964187,
    "sum_squared": 0.4781170077939424,
    "min": 0.6914600550964187,
    "max": 0.6914600550964187,
    "mean": 0.6914600550964187,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7630924630924631,
    "sum_squared": 0.5823101072285223,
    "min": 0.7630924630924631,
    "max": 0.7630924630924631,
    "mean": 0.7630924630924631,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7630924630924631,
    "sum_squared": 0.5823101072285223,
    "min": 0.7630924630924631,
    "max": 0.7630924630924631,
    "mean": 0.7630924630924631,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.47949041914142077,
    "sum_squared": 0.22991106204841535,
    "min": 0.47949041914142077,
    "max": 0.47949041914142077,
    "mean": 0.47949041914142077,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.47949041914142077,
    "sum_squared": 0.22991106204841535,
    "min": 0.47949041914142077,
    "max": 0.47949041914142077,
    "mean": 0.47949041914142077,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.13882378668947293,
    "sum_squared": 0.019272043750804282,
    "min": 0.13882378668947293,
    "max": 0.13882378668947293,
    "mean": 0.13882378668947293,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.13882378668947293,
    "sum_squared": 0.019272043750804282,
    "min": 0.13882378668947293,
    "max": 0.13882378668947293,
    "mean": 0.13882378668947293,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "logprob",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "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": "logprob",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "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": "num_perplexity_tokens",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 5.2727272727272725,
    "sum_squared": 27.801652892561982,
    "min": 5.2727272727272725,
    "max": 5.2727272727272725,
    "mean": 5.2727272727272725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_perplexity_tokens",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 5.2727272727272725,
    "sum_squared": 27.801652892561982,
    "min": 5.2727272727272725,
    "max": 5.2727272727272725,
    "mean": 5.2727272727272725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 20.454545454545453,
    "sum_squared": 418.3884297520661,
    "min": 20.454545454545453,
    "max": 20.454545454545453,
    "mean": 20.454545454545453,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 20.454545454545453,
    "sum_squared": 418.3884297520661,
    "min": 20.454545454545453,
    "max": 20.454545454545453,
    "mean": 20.454545454545453,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_instances",
      "split": "test"
    },
    "count": 1,
    "sum": 39.0,
    "sum_squared": 1521.0,
    "min": 39.0,
    "max": 39.0,
    "mean": 39.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_instances",
      "split": "valid"
    },
    "count": 1,
    "sum": 11.0,
    "sum_squared": 121.0,
    "min": 11.0,
    "max": 11.0,
    "mean": 11.0,
    "variance": 0.0,
    "stddev": 0.0
  }
]