[
  {
    "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": 3638.5641025641025,
    "sum_squared": 13239148.728468113,
    "min": 3638.5641025641025,
    "max": 3638.5641025641025,
    "mean": 3638.5641025641025,
    "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": 0.8982290365757086,
    "sum_squared": 0.8068154021477257,
    "min": 0.8982290365757086,
    "max": 0.8982290365757086,
    "mean": 0.8982290365757086,
    "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": 4.538461538461538,
    "sum_squared": 20.59763313609467,
    "min": 4.538461538461538,
    "max": 4.538461538461538,
    "mean": 4.538461538461538,
    "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": 0.28205128205128205,
    "sum_squared": 0.07955292570677186,
    "min": 0.28205128205128205,
    "max": 0.28205128205128205,
    "mean": 0.28205128205128205,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "test"
    },
    "count": 1,
    "sum": 0.6153846153846154,
    "sum_squared": 0.37869822485207105,
    "min": 0.6153846153846154,
    "max": 0.6153846153846154,
    "mean": 0.6153846153846154,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "test"
    },
    "count": 1,
    "sum": 0.8092498969017188,
    "sum_squared": 0.6548853956354425,
    "min": 0.8092498969017188,
    "max": 0.8092498969017188,
    "mean": 0.8092498969017188,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "test"
    },
    "count": 1,
    "sum": 0.8023182230874539,
    "sum_squared": 0.6437145310982095,
    "min": 0.8023182230874539,
    "max": 0.8023182230874539,
    "mean": 0.8023182230874539,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "test"
    },
    "count": 1,
    "sum": 0.5658146421334544,
    "sum_squared": 0.3201462092526091,
    "min": 0.5658146421334544,
    "max": 0.5658146421334544,
    "mean": 0.5658146421334544,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "test"
    },
    "count": 1,
    "sum": 0.07506891129185964,
    "sum_squared": 0.005635341442545091,
    "min": 0.07506891129185964,
    "max": 0.07506891129185964,
    "mean": 0.07506891129185964,
    "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": 20.794871794871796,
    "sum_squared": 432.42669296515453,
    "min": 20.794871794871796,
    "max": 20.794871794871796,
    "mean": 20.794871794871796,
    "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": 3667.3636363636365,
    "sum_squared": 13449556.041322315,
    "min": 3667.3636363636365,
    "max": 3667.3636363636365,
    "mean": 3667.3636363636365,
    "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": 0.9675865606828169,
    "sum_squared": 0.9362237524140025,
    "min": 0.9675865606828169,
    "max": 0.9675865606828169,
    "mean": 0.9675865606828169,
    "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": 4.636363636363637,
    "sum_squared": 21.495867768595044,
    "min": 4.636363636363637,
    "max": 4.636363636363637,
    "mean": 4.636363636363637,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prompt_truncated",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "max_prob",
      "split": "valid"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.18181818181818182,
    "sum_squared": 0.03305785123966942,
    "min": 0.18181818181818182,
    "max": 0.18181818181818182,
    "mean": 0.18181818181818182,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.5454545454545454,
    "sum_squared": 0.29752066115702475,
    "min": 0.5454545454545454,
    "max": 0.5454545454545454,
    "mean": 0.5454545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "f1_score",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.778236914600551,
    "sum_squared": 0.6056526952469853,
    "min": 0.778236914600551,
    "max": 0.778236914600551,
    "mean": 0.778236914600551,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "rouge_l",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.8305250305250305,
    "sum_squared": 0.6897718263286028,
    "min": 0.8305250305250305,
    "max": 0.8305250305250305,
    "mean": 0.8305250305250305,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_1",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.5475660393344957,
    "sum_squared": 0.29982856743246644,
    "min": 0.5475660393344957,
    "max": 0.5475660393344957,
    "mean": 0.5475660393344957,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "bleu_4",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.1221019939089528,
    "sum_squared": 0.014908896916541944,
    "min": 0.1221019939089528,
    "max": 0.1221019939089528,
    "mean": 0.1221019939089528,
    "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": 18.0,
    "sum_squared": 324.0,
    "min": 18.0,
    "max": 18.0,
    "mean": 18.0,
    "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": 3638.5641025641025,
    "sum_squared": 13239148.728468113,
    "min": 3638.5641025641025,
    "max": 3638.5641025641025,
    "mean": 3638.5641025641025,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_prompt_tokens",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 3638.5641025641025,
    "sum_squared": 13239148.728468113,
    "min": 3638.5641025641025,
    "max": 3638.5641025641025,
    "mean": 3638.5641025641025,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 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": 0.8982290365757086,
    "sum_squared": 0.8068154021477257,
    "min": 0.8982290365757086,
    "max": 0.8982290365757086,
    "mean": 0.8982290365757086,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8982290365757086,
    "sum_squared": 0.8068154021477257,
    "min": 0.8982290365757086,
    "max": 0.8982290365757086,
    "mean": 0.8982290365757086,
    "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": 4.538461538461538,
    "sum_squared": 20.59763313609467,
    "min": 4.538461538461538,
    "max": 4.538461538461538,
    "mean": 4.538461538461538,
    "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": 4.538461538461538,
    "sum_squared": 20.59763313609467,
    "min": 4.538461538461538,
    "max": 4.538461538461538,
    "mean": 4.538461538461538,
    "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": 0.28205128205128205,
    "sum_squared": 0.07955292570677186,
    "min": 0.28205128205128205,
    "max": 0.28205128205128205,
    "mean": 0.28205128205128205,
    "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": 0.28205128205128205,
    "sum_squared": 0.07955292570677186,
    "min": 0.28205128205128205,
    "max": 0.28205128205128205,
    "mean": 0.28205128205128205,
    "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": 0.6153846153846154,
    "sum_squared": 0.37869822485207105,
    "min": 0.6153846153846154,
    "max": 0.6153846153846154,
    "mean": 0.6153846153846154,
    "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": 0.6153846153846154,
    "sum_squared": 0.37869822485207105,
    "min": 0.6153846153846154,
    "max": 0.6153846153846154,
    "mean": 0.6153846153846154,
    "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": 0.8092498969017188,
    "sum_squared": 0.6548853956354425,
    "min": 0.8092498969017188,
    "max": 0.8092498969017188,
    "mean": 0.8092498969017188,
    "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": 0.8092498969017188,
    "sum_squared": 0.6548853956354425,
    "min": 0.8092498969017188,
    "max": 0.8092498969017188,
    "mean": 0.8092498969017188,
    "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": 0.8023182230874539,
    "sum_squared": 0.6437145310982095,
    "min": 0.8023182230874539,
    "max": 0.8023182230874539,
    "mean": 0.8023182230874539,
    "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": 0.8023182230874539,
    "sum_squared": 0.6437145310982095,
    "min": 0.8023182230874539,
    "max": 0.8023182230874539,
    "mean": 0.8023182230874539,
    "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": 0.5658146421334544,
    "sum_squared": 0.3201462092526091,
    "min": 0.5658146421334544,
    "max": 0.5658146421334544,
    "mean": 0.5658146421334544,
    "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": 0.5658146421334544,
    "sum_squared": 0.3201462092526091,
    "min": 0.5658146421334544,
    "max": 0.5658146421334544,
    "mean": 0.5658146421334544,
    "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": 0.07506891129185964,
    "sum_squared": 0.005635341442545091,
    "min": 0.07506891129185964,
    "max": 0.07506891129185964,
    "mean": 0.07506891129185964,
    "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": 0.07506891129185964,
    "sum_squared": 0.005635341442545091,
    "min": 0.07506891129185964,
    "max": 0.07506891129185964,
    "mean": 0.07506891129185964,
    "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": 20.794871794871796,
    "sum_squared": 432.42669296515453,
    "min": 20.794871794871796,
    "max": 20.794871794871796,
    "mean": 20.794871794871796,
    "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": 20.794871794871796,
    "sum_squared": 432.42669296515453,
    "min": 20.794871794871796,
    "max": 20.794871794871796,
    "mean": 20.794871794871796,
    "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": 3667.3636363636365,
    "sum_squared": 13449556.041322315,
    "min": 3667.3636363636365,
    "max": 3667.3636363636365,
    "mean": 3667.3636363636365,
    "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": 3667.3636363636365,
    "sum_squared": 13449556.041322315,
    "min": 3667.3636363636365,
    "max": 3667.3636363636365,
    "mean": 3667.3636363636365,
    "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": 0.9675865606828169,
    "sum_squared": 0.9362237524140025,
    "min": 0.9675865606828169,
    "max": 0.9675865606828169,
    "mean": 0.9675865606828169,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.9675865606828169,
    "sum_squared": 0.9362237524140025,
    "min": 0.9675865606828169,
    "max": 0.9675865606828169,
    "mean": 0.9675865606828169,
    "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": 4.636363636363637,
    "sum_squared": 21.495867768595044,
    "min": 4.636363636363637,
    "max": 4.636363636363637,
    "mean": 4.636363636363637,
    "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": 4.636363636363637,
    "sum_squared": 21.495867768595044,
    "min": 4.636363636363637,
    "max": 4.636363636363637,
    "mean": 4.636363636363637,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prompt_truncated",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prompt_truncated",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.0,
    "sum_squared": 0.0,
    "min": 0.0,
    "max": 0.0,
    "mean": 0.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "max_prob",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "max_prob",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.18181818181818182,
    "sum_squared": 0.03305785123966942,
    "min": 0.18181818181818182,
    "max": 0.18181818181818182,
    "mean": 0.18181818181818182,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.18181818181818182,
    "sum_squared": 0.03305785123966942,
    "min": 0.18181818181818182,
    "max": 0.18181818181818182,
    "mean": 0.18181818181818182,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.5454545454545454,
    "sum_squared": 0.29752066115702475,
    "min": 0.5454545454545454,
    "max": 0.5454545454545454,
    "mean": 0.5454545454545454,
    "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": 0.5454545454545454,
    "sum_squared": 0.29752066115702475,
    "min": 0.5454545454545454,
    "max": 0.5454545454545454,
    "mean": 0.5454545454545454,
    "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": 0.778236914600551,
    "sum_squared": 0.6056526952469853,
    "min": 0.778236914600551,
    "max": 0.778236914600551,
    "mean": 0.778236914600551,
    "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": 0.778236914600551,
    "sum_squared": 0.6056526952469853,
    "min": 0.778236914600551,
    "max": 0.778236914600551,
    "mean": 0.778236914600551,
    "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": 0.8305250305250305,
    "sum_squared": 0.6897718263286028,
    "min": 0.8305250305250305,
    "max": 0.8305250305250305,
    "mean": 0.8305250305250305,
    "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": 0.8305250305250305,
    "sum_squared": 0.6897718263286028,
    "min": 0.8305250305250305,
    "max": 0.8305250305250305,
    "mean": 0.8305250305250305,
    "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": 0.5475660393344957,
    "sum_squared": 0.29982856743246644,
    "min": 0.5475660393344957,
    "max": 0.5475660393344957,
    "mean": 0.5475660393344957,
    "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": 0.5475660393344957,
    "sum_squared": 0.29982856743246644,
    "min": 0.5475660393344957,
    "max": 0.5475660393344957,
    "mean": 0.5475660393344957,
    "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": 0.1221019939089528,
    "sum_squared": 0.014908896916541944,
    "min": 0.1221019939089528,
    "max": 0.1221019939089528,
    "mean": 0.1221019939089528,
    "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": 0.1221019939089528,
    "sum_squared": 0.014908896916541944,
    "min": 0.1221019939089528,
    "max": 0.1221019939089528,
    "mean": 0.1221019939089528,
    "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": 18.0,
    "sum_squared": 324.0,
    "min": 18.0,
    "max": 18.0,
    "mean": 18.0,
    "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": 18.0,
    "sum_squared": 324.0,
    "min": 18.0,
    "max": 18.0,
    "mean": 18.0,
    "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
  }
]