[
  {
    "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": 3647.74358974359,
    "sum_squared": 13306033.296515452,
    "min": 3647.74358974359,
    "max": 3647.74358974359,
    "mean": 3647.74358974359,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "test"
    },
    "count": 1,
    "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_output_tokens",
      "split": "test"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "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": 1.3853136331607134,
    "sum_squared": 1.9190938622209357,
    "min": 1.3853136331607134,
    "max": 1.3853136331607134,
    "mean": 1.3853136331607134,
    "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": 4.871794871794871,
    "sum_squared": 23.734385272846808,
    "min": 4.871794871794871,
    "max": 4.871794871794871,
    "mean": 4.871794871794871,
    "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": 0.3076923076923077,
    "sum_squared": 0.09467455621301776,
    "min": 0.3076923076923077,
    "max": 0.3076923076923077,
    "mean": 0.3076923076923077,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "test"
    },
    "count": 1,
    "sum": 0.48717948717948717,
    "sum_squared": 0.2373438527284681,
    "min": 0.48717948717948717,
    "max": 0.48717948717948717,
    "mean": 0.48717948717948717,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "test"
    },
    "count": 1,
    "sum": 0.6709499615429599,
    "sum_squared": 0.4501738508944994,
    "min": 0.6709499615429599,
    "max": 0.6709499615429599,
    "mean": 0.6709499615429599,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "test"
    },
    "count": 1,
    "sum": 0.691480014238635,
    "sum_squared": 0.4781446100914629,
    "min": 0.691480014238635,
    "max": 0.691480014238635,
    "mean": 0.691480014238635,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "test"
    },
    "count": 1,
    "sum": 0.5407203931382368,
    "sum_squared": 0.2923785435555694,
    "min": 0.5407203931382368,
    "max": 0.5407203931382368,
    "mean": 0.5407203931382368,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "test"
    },
    "count": 1,
    "sum": 0.08205128205128205,
    "sum_squared": 0.00673241288625904,
    "min": 0.08205128205128205,
    "max": 0.08205128205128205,
    "mean": 0.08205128205128205,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "test"
    },
    "count": 1,
    "sum": 23.487179487179485,
    "sum_squared": 551.6476002629848,
    "min": 23.487179487179485,
    "max": 23.487179487179485,
    "mean": 23.487179487179485,
    "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": 3638.181818181818,
    "sum_squared": 13236366.94214876,
    "min": 3638.181818181818,
    "max": 3638.181818181818,
    "mean": 3638.181818181818,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "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_output_tokens",
      "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": "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": 1.394901535727761,
    "sum_squared": 1.9457502943756662,
    "min": 1.394901535727761,
    "max": 1.394901535727761,
    "mean": 1.394901535727761,
    "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": 4.909090909090909,
    "sum_squared": 24.09917355371901,
    "min": 4.909090909090909,
    "max": 4.909090909090909,
    "mean": 4.909090909090909,
    "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": 0.09090909090909091,
    "sum_squared": 0.008264462809917356,
    "min": 0.09090909090909091,
    "max": 0.09090909090909091,
    "mean": 0.09090909090909091,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.45454545454545453,
    "sum_squared": 0.20661157024793386,
    "min": 0.45454545454545453,
    "max": 0.45454545454545453,
    "mean": 0.45454545454545453,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.7487313324633899,
    "sum_squared": 0.5605986082124033,
    "min": 0.7487313324633899,
    "max": 0.7487313324633899,
    "mean": 0.7487313324633899,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.829689608636977,
    "sum_squared": 0.6883848466801802,
    "min": 0.829689608636977,
    "max": 0.829689608636977,
    "mean": 0.829689608636977,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.5569414914948082,
    "sum_squared": 0.31018382494846153,
    "min": 0.5569414914948082,
    "max": 0.5569414914948082,
    "mean": 0.5569414914948082,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.15467159893503413,
    "sum_squared": 0.02392330351712005,
    "min": 0.15467159893503413,
    "max": 0.15467159893503413,
    "mean": 0.15467159893503413,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "valid"
    },
    "count": 1,
    "sum": 20.272727272727273,
    "sum_squared": 410.9834710743802,
    "min": 20.272727272727273,
    "max": 20.272727272727273,
    "mean": 20.272727272727273,
    "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": 3647.74358974359,
    "sum_squared": 13306033.296515452,
    "min": 3647.74358974359,
    "max": 3647.74358974359,
    "mean": 3647.74358974359,
    "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": 3647.74358974359,
    "sum_squared": 13306033.296515452,
    "min": 3647.74358974359,
    "max": 3647.74358974359,
    "mean": 3647.74358974359,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.3853136331607134,
    "sum_squared": 1.9190938622209357,
    "min": 1.3853136331607134,
    "max": 1.3853136331607134,
    "mean": 1.3853136331607134,
    "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": 1.3853136331607134,
    "sum_squared": 1.9190938622209357,
    "min": 1.3853136331607134,
    "max": 1.3853136331607134,
    "mean": 1.3853136331607134,
    "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": 4.871794871794871,
    "sum_squared": 23.734385272846808,
    "min": 4.871794871794871,
    "max": 4.871794871794871,
    "mean": 4.871794871794871,
    "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": 4.871794871794871,
    "sum_squared": 23.734385272846808,
    "min": 4.871794871794871,
    "max": 4.871794871794871,
    "mean": 4.871794871794871,
    "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": 0.3076923076923077,
    "sum_squared": 0.09467455621301776,
    "min": 0.3076923076923077,
    "max": 0.3076923076923077,
    "mean": 0.3076923076923077,
    "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": 0.3076923076923077,
    "sum_squared": 0.09467455621301776,
    "min": 0.3076923076923077,
    "max": 0.3076923076923077,
    "mean": 0.3076923076923077,
    "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": 0.48717948717948717,
    "sum_squared": 0.2373438527284681,
    "min": 0.48717948717948717,
    "max": 0.48717948717948717,
    "mean": 0.48717948717948717,
    "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": 0.48717948717948717,
    "sum_squared": 0.2373438527284681,
    "min": 0.48717948717948717,
    "max": 0.48717948717948717,
    "mean": 0.48717948717948717,
    "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": 0.6709499615429599,
    "sum_squared": 0.4501738508944994,
    "min": 0.6709499615429599,
    "max": 0.6709499615429599,
    "mean": 0.6709499615429599,
    "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": 0.6709499615429599,
    "sum_squared": 0.4501738508944994,
    "min": 0.6709499615429599,
    "max": 0.6709499615429599,
    "mean": 0.6709499615429599,
    "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": 0.691480014238635,
    "sum_squared": 0.4781446100914629,
    "min": 0.691480014238635,
    "max": 0.691480014238635,
    "mean": 0.691480014238635,
    "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": 0.691480014238635,
    "sum_squared": 0.4781446100914629,
    "min": 0.691480014238635,
    "max": 0.691480014238635,
    "mean": 0.691480014238635,
    "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": 0.5407203931382368,
    "sum_squared": 0.2923785435555694,
    "min": 0.5407203931382368,
    "max": 0.5407203931382368,
    "mean": 0.5407203931382368,
    "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": 0.5407203931382368,
    "sum_squared": 0.2923785435555694,
    "min": 0.5407203931382368,
    "max": 0.5407203931382368,
    "mean": 0.5407203931382368,
    "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": 0.08205128205128205,
    "sum_squared": 0.00673241288625904,
    "min": 0.08205128205128205,
    "max": 0.08205128205128205,
    "mean": 0.08205128205128205,
    "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": 0.08205128205128205,
    "sum_squared": 0.00673241288625904,
    "min": 0.08205128205128205,
    "max": 0.08205128205128205,
    "mean": 0.08205128205128205,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 23.487179487179485,
    "sum_squared": 551.6476002629848,
    "min": 23.487179487179485,
    "max": 23.487179487179485,
    "mean": 23.487179487179485,
    "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": 23.487179487179485,
    "sum_squared": 551.6476002629848,
    "min": 23.487179487179485,
    "max": 23.487179487179485,
    "mean": 23.487179487179485,
    "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": 3638.181818181818,
    "sum_squared": 13236366.94214876,
    "min": 3638.181818181818,
    "max": 3638.181818181818,
    "mean": 3638.181818181818,
    "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": 3638.181818181818,
    "sum_squared": 13236366.94214876,
    "min": 3638.181818181818,
    "max": 3638.181818181818,
    "mean": 3638.181818181818,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.394901535727761,
    "sum_squared": 1.9457502943756662,
    "min": 1.394901535727761,
    "max": 1.394901535727761,
    "mean": 1.394901535727761,
    "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": 1.394901535727761,
    "sum_squared": 1.9457502943756662,
    "min": 1.394901535727761,
    "max": 1.394901535727761,
    "mean": 1.394901535727761,
    "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": 4.909090909090909,
    "sum_squared": 24.09917355371901,
    "min": 4.909090909090909,
    "max": 4.909090909090909,
    "mean": 4.909090909090909,
    "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": 4.909090909090909,
    "sum_squared": 24.09917355371901,
    "min": 4.909090909090909,
    "max": 4.909090909090909,
    "mean": 4.909090909090909,
    "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": 0.09090909090909091,
    "sum_squared": 0.008264462809917356,
    "min": 0.09090909090909091,
    "max": 0.09090909090909091,
    "mean": 0.09090909090909091,
    "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": 0.09090909090909091,
    "sum_squared": 0.008264462809917356,
    "min": 0.09090909090909091,
    "max": 0.09090909090909091,
    "mean": 0.09090909090909091,
    "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": 0.45454545454545453,
    "sum_squared": 0.20661157024793386,
    "min": 0.45454545454545453,
    "max": 0.45454545454545453,
    "mean": 0.45454545454545453,
    "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": 0.45454545454545453,
    "sum_squared": 0.20661157024793386,
    "min": 0.45454545454545453,
    "max": 0.45454545454545453,
    "mean": 0.45454545454545453,
    "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": 0.7487313324633899,
    "sum_squared": 0.5605986082124033,
    "min": 0.7487313324633899,
    "max": 0.7487313324633899,
    "mean": 0.7487313324633899,
    "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": 0.7487313324633899,
    "sum_squared": 0.5605986082124033,
    "min": 0.7487313324633899,
    "max": 0.7487313324633899,
    "mean": 0.7487313324633899,
    "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": 0.829689608636977,
    "sum_squared": 0.6883848466801802,
    "min": 0.829689608636977,
    "max": 0.829689608636977,
    "mean": 0.829689608636977,
    "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": 0.829689608636977,
    "sum_squared": 0.6883848466801802,
    "min": 0.829689608636977,
    "max": 0.829689608636977,
    "mean": 0.829689608636977,
    "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": 0.5569414914948082,
    "sum_squared": 0.31018382494846153,
    "min": 0.5569414914948082,
    "max": 0.5569414914948082,
    "mean": 0.5569414914948082,
    "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": 0.5569414914948082,
    "sum_squared": 0.31018382494846153,
    "min": 0.5569414914948082,
    "max": 0.5569414914948082,
    "mean": 0.5569414914948082,
    "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": 0.15467159893503413,
    "sum_squared": 0.02392330351712005,
    "min": 0.15467159893503413,
    "max": 0.15467159893503413,
    "mean": 0.15467159893503413,
    "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": 0.15467159893503413,
    "sum_squared": 0.02392330351712005,
    "min": 0.15467159893503413,
    "max": 0.15467159893503413,
    "mean": 0.15467159893503413,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 20.272727272727273,
    "sum_squared": 410.9834710743802,
    "min": 20.272727272727273,
    "max": 20.272727272727273,
    "mean": 20.272727272727273,
    "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": 20.272727272727273,
    "sum_squared": 410.9834710743802,
    "min": 20.272727272727273,
    "max": 20.272727272727273,
    "mean": 20.272727272727273,
    "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
  }
]