{
  "schema_version": "1.1",
  "benchmark_release": "repeated-fold-v2.2",
  "estimator_scope": "same_processed_hepg2_dataset",
  "metrics": {
    "rmse_difference_vs_zero": {
      "estimator": "mean_across_repeated_balanced_fold_realizations",
      "mean": -0.006291882263806958,
      "ci95": [
        -0.007352181034921496,
        -0.005240325481195381
      ],
      "n_realizations": 20,
      "targets_per_realization": 2160,
      "bootstrap_replicates": 10000,
      "bootstrap_seed": 20260737,
      "source_object": "api/v1/v22_repeat_joint_uncertainty.json",
      "source_table": "downloads/v22_repeat_joint_uncertainty.tsv"
    },
    "rmse_difference_vs_train_mean": {
      "estimator": "mean_across_repeated_balanced_fold_realizations",
      "mean": -0.0004662719641432719,
      "ci95": [
        -0.0007022066115125746,
        -0.00022487605050965128
      ],
      "n_realizations": 20,
      "targets_per_realization": 2160,
      "bootstrap_replicates": 10000,
      "bootstrap_seed": 20260737,
      "source_object": "api/v1/v22_repeat_joint_uncertainty.json",
      "source_table": "downloads/v22_repeat_joint_uncertainty.tsv"
    },
    "raw_cosine_difference_vs_train_mean": {
      "estimator": "mean_across_repeated_balanced_fold_realizations",
      "mean": 0.0012235986991812387,
      "ci95": [
        -0.00021601119406292822,
        0.0024983961379381153
      ],
      "n_realizations": 20,
      "targets_per_realization": 2160,
      "bootstrap_replicates": 10000,
      "bootstrap_seed": 20260737,
      "source_object": "api/v1/v22_repeat_joint_uncertainty.json",
      "source_table": "downloads/v22_repeat_joint_uncertainty.tsv"
    },
    "perturbation_specific_residual_cosine": {
      "estimator": "mean_across_repeated_balanced_fold_realizations",
      "mean": 0.045280892411227326,
      "ci95": [
        0.02964996627740961,
        0.06117821633174975
      ],
      "n_realizations": 20,
      "targets_per_realization": 2160,
      "bootstrap_replicates": 10000,
      "bootstrap_seed": 20260737,
      "source_object": "api/v1/v22_repeat_joint_uncertainty.json",
      "source_table": "downloads/v22_repeat_joint_uncertainty.tsv"
    }
  },
  "absolute_and_relative_audit": {
    "model_mean_rmse": 0.11750674428366534,
    "train_mean_rmse": 0.11797301624780862,
    "zero_rmse": 0.12379862654747228,
    "relative_improvement_vs_train_mean": 0.003952365435038386,
    "relative_improvement_vs_zero": 0.05082352235462188,
    "target_distribution": {
      "target_count": 2160,
      "improved_vs_train_mean": 1128,
      "worsened_vs_train_mean": 1032,
      "exactly_unchanged_vs_train_mean": 0
    }
  }
}