[
  {
    "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": 3673.6666666666665,
    "sum_squared": 13495826.777777776,
    "min": 3673.6666666666665,
    "max": 3673.6666666666665,
    "mean": 3673.6666666666665,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "test"
    },
    "count": 1,
    "sum": 49.84615384615385,
    "sum_squared": 2484.639053254438,
    "min": 49.84615384615385,
    "max": 49.84615384615385,
    "mean": 49.84615384615385,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "test"
    },
    "count": 1,
    "sum": 49.84615384615385,
    "sum_squared": 2484.639053254438,
    "min": 49.84615384615385,
    "max": 49.84615384615385,
    "mean": 49.84615384615385,
    "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": 5.574288093126738,
    "sum_squared": 31.07268774517452,
    "min": 5.574288093126738,
    "max": 5.574288093126738,
    "mean": 5.574288093126738,
    "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": 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": "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "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": "f1_score",
      "split": "test"
    },
    "count": 1,
    "sum": 0.19303124034683725,
    "sum_squared": 0.03726105974983845,
    "min": 0.19303124034683725,
    "max": 0.19303124034683725,
    "mean": 0.19303124034683725,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "test"
    },
    "count": 1,
    "sum": 0.16609678980171586,
    "sum_squared": 0.02758814358243538,
    "min": 0.16609678980171586,
    "max": 0.16609678980171586,
    "mean": 0.16609678980171586,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "test"
    },
    "count": 1,
    "sum": 0.09320178822615054,
    "sum_squared": 0.008686573328552213,
    "min": 0.09320178822615054,
    "max": 0.09320178822615054,
    "mean": 0.09320178822615054,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "test"
    },
    "count": 1,
    "sum": 0.01426482047348085,
    "sum_squared": 0.0002034851031406384,
    "min": 0.01426482047348085,
    "max": 0.01426482047348085,
    "mean": 0.01426482047348085,
    "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": 49.84615384615385,
    "sum_squared": 2484.639053254438,
    "min": 49.84615384615385,
    "max": 49.84615384615385,
    "mean": 49.84615384615385,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "test"
    },
    "count": 1,
    "sum": 232.2051282051282,
    "sum_squared": 53919.221564760024,
    "min": 232.2051282051282,
    "max": 232.2051282051282,
    "mean": 232.2051282051282,
    "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": 3651.181818181818,
    "sum_squared": 13331128.669421487,
    "min": 3651.181818181818,
    "max": 3651.181818181818,
    "mean": 3651.181818181818,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "valid"
    },
    "count": 1,
    "sum": 60.36363636363637,
    "sum_squared": 3643.768595041323,
    "min": 60.36363636363637,
    "max": 60.36363636363637,
    "mean": 60.36363636363637,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "valid"
    },
    "count": 1,
    "sum": 60.36363636363637,
    "sum_squared": 3643.768595041323,
    "min": 60.36363636363637,
    "max": 60.36363636363637,
    "mean": 60.36363636363637,
    "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": 6.128317919644442,
    "sum_squared": 37.556280524235184,
    "min": 6.128317919644442,
    "max": 6.128317919644442,
    "mean": 6.128317919644442,
    "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "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": "f1_score",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.21031996112390014,
    "sum_squared": 0.044234486047158864,
    "min": 0.21031996112390014,
    "max": 0.21031996112390014,
    "mean": 0.21031996112390014,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.17281140533829328,
    "sum_squared": 0.0298637818149959,
    "min": 0.17281140533829328,
    "max": 0.17281140533829328,
    "mean": 0.17281140533829328,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.0937424902445759,
    "sum_squared": 0.008787654477254407,
    "min": 0.0937424902445759,
    "max": 0.0937424902445759,
    "mean": 0.0937424902445759,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.018912933631934525,
    "sum_squared": 0.0003576990585659601,
    "min": 0.018912933631934525,
    "max": 0.018912933631934525,
    "mean": 0.018912933631934525,
    "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": 60.36363636363637,
    "sum_squared": 3643.768595041323,
    "min": 60.36363636363637,
    "max": 60.36363636363637,
    "mean": 60.36363636363637,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_bytes",
      "split": "valid"
    },
    "count": 1,
    "sum": 279.54545454545456,
    "sum_squared": 78145.6611570248,
    "min": 279.54545454545456,
    "max": 279.54545454545456,
    "mean": 279.54545454545456,
    "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": 3673.6666666666665,
    "sum_squared": 13495826.777777776,
    "min": 3673.6666666666665,
    "max": 3673.6666666666665,
    "mean": 3673.6666666666665,
    "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": 3673.6666666666665,
    "sum_squared": 13495826.777777776,
    "min": 3673.6666666666665,
    "max": 3673.6666666666665,
    "mean": 3673.6666666666665,
    "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": 49.84615384615385,
    "sum_squared": 2484.639053254438,
    "min": 49.84615384615385,
    "max": 49.84615384615385,
    "mean": 49.84615384615385,
    "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": 49.84615384615385,
    "sum_squared": 2484.639053254438,
    "min": 49.84615384615385,
    "max": 49.84615384615385,
    "mean": 49.84615384615385,
    "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": 49.84615384615385,
    "sum_squared": 2484.639053254438,
    "min": 49.84615384615385,
    "max": 49.84615384615385,
    "mean": 49.84615384615385,
    "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": 49.84615384615385,
    "sum_squared": 2484.639053254438,
    "min": 49.84615384615385,
    "max": 49.84615384615385,
    "mean": 49.84615384615385,
    "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": 5.574288093126738,
    "sum_squared": 31.07268774517452,
    "min": 5.574288093126738,
    "max": 5.574288093126738,
    "mean": 5.574288093126738,
    "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": 5.574288093126738,
    "sum_squared": 31.07268774517452,
    "min": 5.574288093126738,
    "max": 5.574288093126738,
    "mean": 5.574288093126738,
    "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": 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": "test",
      "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": "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "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": 0.19303124034683725,
    "sum_squared": 0.03726105974983845,
    "min": 0.19303124034683725,
    "max": 0.19303124034683725,
    "mean": 0.19303124034683725,
    "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": 0.19303124034683725,
    "sum_squared": 0.03726105974983845,
    "min": 0.19303124034683725,
    "max": 0.19303124034683725,
    "mean": 0.19303124034683725,
    "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": 0.16609678980171586,
    "sum_squared": 0.02758814358243538,
    "min": 0.16609678980171586,
    "max": 0.16609678980171586,
    "mean": 0.16609678980171586,
    "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": 0.16609678980171586,
    "sum_squared": 0.02758814358243538,
    "min": 0.16609678980171586,
    "max": 0.16609678980171586,
    "mean": 0.16609678980171586,
    "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": 0.09320178822615054,
    "sum_squared": 0.008686573328552213,
    "min": 0.09320178822615054,
    "max": 0.09320178822615054,
    "mean": 0.09320178822615054,
    "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": 0.09320178822615054,
    "sum_squared": 0.008686573328552213,
    "min": 0.09320178822615054,
    "max": 0.09320178822615054,
    "mean": 0.09320178822615054,
    "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": 0.01426482047348085,
    "sum_squared": 0.0002034851031406384,
    "min": 0.01426482047348085,
    "max": 0.01426482047348085,
    "mean": 0.01426482047348085,
    "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": 0.01426482047348085,
    "sum_squared": 0.0002034851031406384,
    "min": 0.01426482047348085,
    "max": 0.01426482047348085,
    "mean": 0.01426482047348085,
    "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": 49.84615384615385,
    "sum_squared": 2484.639053254438,
    "min": 49.84615384615385,
    "max": 49.84615384615385,
    "mean": 49.84615384615385,
    "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": 49.84615384615385,
    "sum_squared": 2484.639053254438,
    "min": 49.84615384615385,
    "max": 49.84615384615385,
    "mean": 49.84615384615385,
    "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": 232.2051282051282,
    "sum_squared": 53919.221564760024,
    "min": 232.2051282051282,
    "max": 232.2051282051282,
    "mean": 232.2051282051282,
    "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": 232.2051282051282,
    "sum_squared": 53919.221564760024,
    "min": 232.2051282051282,
    "max": 232.2051282051282,
    "mean": 232.2051282051282,
    "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": 3651.181818181818,
    "sum_squared": 13331128.669421487,
    "min": 3651.181818181818,
    "max": 3651.181818181818,
    "mean": 3651.181818181818,
    "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": 3651.181818181818,
    "sum_squared": 13331128.669421487,
    "min": 3651.181818181818,
    "max": 3651.181818181818,
    "mean": 3651.181818181818,
    "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": 60.36363636363637,
    "sum_squared": 3643.768595041323,
    "min": 60.36363636363637,
    "max": 60.36363636363637,
    "mean": 60.36363636363637,
    "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": 60.36363636363637,
    "sum_squared": 3643.768595041323,
    "min": 60.36363636363637,
    "max": 60.36363636363637,
    "mean": 60.36363636363637,
    "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": 60.36363636363637,
    "sum_squared": 3643.768595041323,
    "min": 60.36363636363637,
    "max": 60.36363636363637,
    "mean": 60.36363636363637,
    "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": 60.36363636363637,
    "sum_squared": 3643.768595041323,
    "min": 60.36363636363637,
    "max": 60.36363636363637,
    "mean": 60.36363636363637,
    "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": 6.128317919644442,
    "sum_squared": 37.556280524235184,
    "min": 6.128317919644442,
    "max": 6.128317919644442,
    "mean": 6.128317919644442,
    "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": 6.128317919644442,
    "sum_squared": 37.556280524235184,
    "min": 6.128317919644442,
    "max": 6.128317919644442,
    "mean": 6.128317919644442,
    "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "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": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "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": 0.21031996112390014,
    "sum_squared": 0.044234486047158864,
    "min": 0.21031996112390014,
    "max": 0.21031996112390014,
    "mean": 0.21031996112390014,
    "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": 0.21031996112390014,
    "sum_squared": 0.044234486047158864,
    "min": 0.21031996112390014,
    "max": 0.21031996112390014,
    "mean": 0.21031996112390014,
    "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": 0.17281140533829328,
    "sum_squared": 0.0298637818149959,
    "min": 0.17281140533829328,
    "max": 0.17281140533829328,
    "mean": 0.17281140533829328,
    "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": 0.17281140533829328,
    "sum_squared": 0.0298637818149959,
    "min": 0.17281140533829328,
    "max": 0.17281140533829328,
    "mean": 0.17281140533829328,
    "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": 0.0937424902445759,
    "sum_squared": 0.008787654477254407,
    "min": 0.0937424902445759,
    "max": 0.0937424902445759,
    "mean": 0.0937424902445759,
    "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": 0.0937424902445759,
    "sum_squared": 0.008787654477254407,
    "min": 0.0937424902445759,
    "max": 0.0937424902445759,
    "mean": 0.0937424902445759,
    "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": 0.018912933631934525,
    "sum_squared": 0.0003576990585659601,
    "min": 0.018912933631934525,
    "max": 0.018912933631934525,
    "mean": 0.018912933631934525,
    "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": 0.018912933631934525,
    "sum_squared": 0.0003576990585659601,
    "min": 0.018912933631934525,
    "max": 0.018912933631934525,
    "mean": 0.018912933631934525,
    "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": 60.36363636363637,
    "sum_squared": 3643.768595041323,
    "min": 60.36363636363637,
    "max": 60.36363636363637,
    "mean": 60.36363636363637,
    "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": 60.36363636363637,
    "sum_squared": 3643.768595041323,
    "min": 60.36363636363637,
    "max": 60.36363636363637,
    "mean": 60.36363636363637,
    "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": 279.54545454545456,
    "sum_squared": 78145.6611570248,
    "min": 279.54545454545456,
    "max": 279.54545454545456,
    "mean": 279.54545454545456,
    "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": 279.54545454545456,
    "sum_squared": 78145.6611570248,
    "min": 279.54545454545456,
    "max": 279.54545454545456,
    "mean": 279.54545454545456,
    "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
  }
]