[
  {
    "name": {
      "name": "num_references",
      "split": "test"
    },
    "count": 1,
    "sum": 4.0,
    "sum_squared": 16.0,
    "min": 4.0,
    "max": 4.0,
    "mean": 4.0,
    "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": 493.95454545454544,
    "sum_squared": 243991.0929752066,
    "min": 493.95454545454544,
    "max": 493.95454545454544,
    "mean": 493.95454545454544,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "test"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "test"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "training_co2_cost",
      "split": "test"
    },
    "count": 0,
    "sum": 0,
    "sum_squared": 0
  },
  {
    "name": {
      "name": "training_energy_cost",
      "split": "test"
    },
    "count": 0,
    "sum": 0,
    "sum_squared": 0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "test"
    },
    "count": 1,
    "sum": 1.4668538516218013,
    "sum_squared": 2.1516602220177132,
    "min": 1.4668538516218013,
    "max": 1.4668538516218013,
    "mean": 1.4668538516218013,
    "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": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match@5",
      "split": "test"
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "test"
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match@5",
      "split": "test"
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match",
      "split": "test"
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match@5",
      "split": "test"
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match",
      "split": "test"
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match@5",
      "split": "test"
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_references",
      "split": "valid"
    },
    "count": 1,
    "sum": 4.0,
    "sum_squared": 16.0,
    "min": 4.0,
    "max": 4.0,
    "mean": 4.0,
    "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": 513.1666666666666,
    "sum_squared": 263340.02777777775,
    "min": 513.1666666666666,
    "max": 513.1666666666666,
    "mean": 513.1666666666666,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "valid"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "valid"
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "training_co2_cost",
      "split": "valid"
    },
    "count": 0,
    "sum": 0,
    "sum_squared": 0
  },
  {
    "name": {
      "name": "training_energy_cost",
      "split": "valid"
    },
    "count": 0,
    "sum": 0,
    "sum_squared": 0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "valid"
    },
    "count": 1,
    "sum": 1.6378906567891438,
    "sum_squared": 2.682685803597173,
    "min": 1.6378906567891438,
    "max": 1.6378906567891438,
    "mean": 1.6378906567891438,
    "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": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match@5",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match@5",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match@5",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match@5",
      "split": "valid"
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "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": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "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": 4.0,
    "sum_squared": 16.0,
    "min": 4.0,
    "max": 4.0,
    "mean": 4.0,
    "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": 4.0,
    "sum_squared": 16.0,
    "min": 4.0,
    "max": 4.0,
    "mean": 4.0,
    "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": 493.95454545454544,
    "sum_squared": 243991.0929752066,
    "min": 493.95454545454544,
    "max": 493.95454545454544,
    "mean": 493.95454545454544,
    "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": 493.95454545454544,
    "sum_squared": 243991.0929752066,
    "min": 493.95454545454544,
    "max": 493.95454545454544,
    "mean": 493.95454545454544,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.4668538516218013,
    "sum_squared": 2.1516602220177132,
    "min": 1.4668538516218013,
    "max": 1.4668538516218013,
    "mean": 1.4668538516218013,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.4668538516218013,
    "sum_squared": 2.1516602220177132,
    "min": 1.4668538516218013,
    "max": 1.4668538516218013,
    "mean": 1.4668538516218013,
    "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": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "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": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match@5",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match@5",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "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": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "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": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match@5",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match@5",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match@5",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match@5",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match@5",
      "split": "test",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match@5",
      "split": "test",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.7954545454545454,
    "sum_squared": 0.6327479338842975,
    "min": 0.7954545454545454,
    "max": 0.7954545454545454,
    "mean": 0.7954545454545454,
    "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": 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": "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_references",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 4.0,
    "sum_squared": 16.0,
    "min": 4.0,
    "max": 4.0,
    "mean": 4.0,
    "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": 4.0,
    "sum_squared": 16.0,
    "min": 4.0,
    "max": 4.0,
    "mean": 4.0,
    "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": 513.1666666666666,
    "sum_squared": 263340.02777777775,
    "min": 513.1666666666666,
    "max": 513.1666666666666,
    "mean": 513.1666666666666,
    "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": 513.1666666666666,
    "sum_squared": 263340.02777777775,
    "min": 513.1666666666666,
    "max": 513.1666666666666,
    "mean": 513.1666666666666,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_completion_tokens",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_output_tokens",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.0,
    "sum_squared": 1.0,
    "min": 1.0,
    "max": 1.0,
    "mean": 1.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.6378906567891438,
    "sum_squared": 2.682685803597173,
    "min": 1.6378906567891438,
    "max": 1.6378906567891438,
    "mean": 1.6378906567891438,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "inference_runtime",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 1.6378906567891438,
    "sum_squared": 2.682685803597173,
    "min": 1.6378906567891438,
    "max": 1.6378906567891438,
    "mean": 1.6378906567891438,
    "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": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "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": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match@5",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "exact_match@5",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match@5",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_exact_match@5",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match@5",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "prefix_exact_match@5",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match@5",
      "split": "valid",
      "perturbation": {
        "name": "robustness",
        "robustness": true,
        "fairness": false,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "quasi_prefix_exact_match@5",
      "split": "valid",
      "perturbation": {
        "name": "fairness",
        "robustness": false,
        "fairness": true,
        "computed_on": "worst"
      }
    },
    "count": 1,
    "sum": 0.8333333333333334,
    "sum_squared": 0.6944444444444445,
    "min": 0.8333333333333334,
    "max": 0.8333333333333334,
    "mean": 0.8333333333333334,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "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": 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": "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_instances",
      "split": "test"
    },
    "count": 1,
    "sum": 44.0,
    "sum_squared": 1936.0,
    "min": 44.0,
    "max": 44.0,
    "mean": 44.0,
    "variance": 0.0,
    "stddev": 0.0
  },
  {
    "name": {
      "name": "num_instances",
      "split": "valid"
    },
    "count": 1,
    "sum": 6.0,
    "sum_squared": 36.0,
    "min": 6.0,
    "max": 6.0,
    "mean": 6.0,
    "variance": 0.0,
    "stddev": 0.0
  }
]