{
  "metrics": {
    "uniform": {
      "rows": 1000,
      "log_loss": 1.0986122886681096,
      "accuracy": 0.341,
      "macro_f1": 0.1695252299279145,
      "brier_sum_three_classes": 0.6666666666666666,
      "per_class": {
        "a_wins": {
          "precision": 0.341,
          "recall": 1.0,
          "f1-score": 0.5085756897837435,
          "support": 341.0
        },
        "b_wins": {
          "precision": 0.0,
          "recall": 0.0,
          "f1-score": 0.0,
          "support": 364.0
        },
        "tie": {
          "precision": 0.0,
          "recall": 0.0,
          "f1-score": 0.0,
          "support": 295.0
        },
        "accuracy": 0.341,
        "macro avg": {
          "precision": 0.11366666666666668,
          "recall": 0.3333333333333333,
          "f1-score": 0.1695252299279145,
          "support": 1000.0
        },
        "weighted avg": {
          "precision": 0.11628100000000001,
          "recall": 0.341,
          "f1-score": 0.1734243102162565,
          "support": 1000.0
        }
      },
      "prediction_counts": [
        1000,
        0,
        0
      ]
    },
    "train_prior": {
      "rows": 1000,
      "log_loss": 1.0953927535292982,
      "accuracy": 0.364,
      "macro_f1": 0.17790811339198434,
      "brier_sum_three_classes": 0.66455184375,
      "per_class": {
        "a_wins": {
          "precision": 0.0,
          "recall": 0.0,
          "f1-score": 0.0,
          "support": 341.0
        },
        "b_wins": {
          "precision": 0.364,
          "recall": 1.0,
          "f1-score": 0.533724340175953,
          "support": 364.0
        },
        "tie": {
          "precision": 0.0,
          "recall": 0.0,
          "f1-score": 0.0,
          "support": 295.0
        },
        "accuracy": 0.364,
        "macro avg": {
          "precision": 0.12133333333333333,
          "recall": 0.3333333333333333,
          "f1-score": 0.17790811339198434,
          "support": 1000.0
        },
        "weighted avg": {
          "precision": 0.132496,
          "recall": 0.364,
          "f1-score": 0.1942756598240469,
          "support": 1000.0
        }
      },
      "prediction_counts": [
        0,
        1000,
        0
      ]
    },
    "single": {
      "rows": 1000,
      "log_loss": 1.080322504043579,
      "accuracy": 0.402,
      "macro_f1": 0.40130730201653936,
      "brier_sum_three_classes": 0.6529792205827425,
      "per_class": {
        "a_wins": {
          "precision": 0.37897310513447435,
          "recall": 0.45454545454545453,
          "f1-score": 0.41333333333333333,
          "support": 341.0
        },
        "b_wins": {
          "precision": 0.45588235294117646,
          "recall": 0.34065934065934067,
          "f1-score": 0.389937106918239,
          "support": 364.0
        },
        "tie": {
          "precision": 0.38557993730407525,
          "recall": 0.41694915254237286,
          "f1-score": 0.4006514657980456,
          "support": 295.0
        },
        "accuracy": 0.402,
        "macro avg": {
          "precision": 0.40681179845990867,
          "recall": 0.4040513159157227,
          "f1-score": 0.40130730201653936,
          "support": 1000.0
        },
        "weighted avg": {
          "precision": 0.4089170868261462,
          "recall": 0.402,
          "f1-score": 0.40107595599532914,
          "support": 1000.0
        }
      },
      "prediction_counts": [
        409,
        272,
        319
      ]
    },
    "swap_average": {
      "rows": 1000,
      "log_loss": 1.071562647819519,
      "accuracy": 0.429,
      "macro_f1": 0.42909672416242434,
      "brier_sum_three_classes": 0.6472357825514672,
      "per_class": {
        "a_wins": {
          "precision": 0.4533762057877814,
          "recall": 0.41348973607038125,
          "f1-score": 0.4325153374233129,
          "support": 341.0
        },
        "b_wins": {
          "precision": 0.48859934853420195,
          "recall": 0.41208791208791207,
          "f1-score": 0.44709388971684055,
          "support": 364.0
        },
        "tie": {
          "precision": 0.3612565445026178,
          "recall": 0.46779661016949153,
          "f1-score": 0.4076809453471196,
          "support": 295.0
        },
        "accuracy": 0.429,
        "macro avg": {
          "precision": 0.4344106996082004,
          "recall": 0.43112475277592827,
          "f1-score": 0.42909672416242434,
          "support": 1000.0
        },
        "weighted avg": {
          "precision": 0.4390221296683552,
          "recall": 0.429,
          "f1-score": 0.4304957847956799,
          "support": 1000.0
        }
      },
      "prediction_counts": [
        311,
        307,
        382
      ]
    }
  },
  "paired_swap_minus_single": {
    "method": "2000 paired connected-group bootstrap replicates, percentile 95% intervals",
    "groups": 889,
    "log_loss_delta": {
      "estimate": -0.008759795688092709,
      "ci95": [
        -0.016371096908631828,
        -0.0013557262337078428
      ]
    },
    "accuracy_delta": {
      "estimate": 0.027,
      "ci95": [
        0.0029319119552615428,
        0.050645808620306286
      ]
    }
  },
  "flip_rate_single": 0.505,
  "mean_aligned_absolute_probability_difference": 0.05893449857831001,
  "swap_equivariance_max_abs": 0.0,
  "slices": {
    "truncated": {
      "uniform": {
        "rows": 547,
        "log_loss": 1.0986122886681098,
        "accuracy": 0.3382084095063985,
        "macro_f1": 0.16848816029143898,
        "brier_sum_three_classes": 0.6666666666666669,
        "per_class": {
          "a_wins": {
            "precision": 0.3382084095063985,
            "recall": 1.0,
            "f1-score": 0.505464480874317,
            "support": 185.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 211.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 151.0
          },
          "accuracy": 0.3382084095063985,
          "macro avg": {
            "precision": 0.11273613650213284,
            "recall": 0.3333333333333333,
            "f1-score": 0.16848816029143898,
            "support": 547.0
          },
          "weighted avg": {
            "precision": 0.11438492826084777,
            "recall": 0.3382084095063985,
            "f1-score": 0.17095233813848013,
            "support": 547.0
          }
        },
        "prediction_counts": [
          547,
          0,
          0
        ]
      },
      "train_prior": {
        "rows": 547,
        "log_loss": 1.092862839052362,
        "accuracy": 0.3857404021937843,
        "macro_f1": 0.18557607739665785,
        "brier_sum_three_classes": 0.6628936321412249,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 185.0
          },
          "b_wins": {
            "precision": 0.3857404021937843,
            "recall": 1.0,
            "f1-score": 0.5567282321899736,
            "support": 211.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 151.0
          },
          "accuracy": 0.3857404021937843,
          "macro avg": {
            "precision": 0.12858013406459476,
            "recall": 0.3333333333333333,
            "f1-score": 0.18557607739665785,
            "support": 547.0
          },
          "weighted avg": {
            "precision": 0.14879565788462248,
            "recall": 0.3857404021937843,
            "f1-score": 0.21475257219759492,
            "support": 547.0
          }
        },
        "prediction_counts": [
          0,
          547,
          0
        ]
      },
      "single": {
        "rows": 547,
        "log_loss": 1.089914321899414,
        "accuracy": 0.376599634369287,
        "macro_f1": 0.36939449972488775,
        "brier_sum_three_classes": 0.6600291493741667,
        "per_class": {
          "a_wins": {
            "precision": 0.34317343173431736,
            "recall": 0.5027027027027027,
            "f1-score": 0.40789473684210525,
            "support": 185.0
          },
          "b_wins": {
            "precision": 0.45098039215686275,
            "recall": 0.32701421800947866,
            "f1-score": 0.3791208791208791,
            "support": 211.0
          },
          "tie": {
            "precision": 0.35772357723577236,
            "recall": 0.2913907284768212,
            "f1-score": 0.32116788321167883,
            "support": 151.0
          },
          "accuracy": 0.376599634369287,
          "macro avg": {
            "precision": 0.3839591337089841,
            "recall": 0.3737025497296675,
            "f1-score": 0.36939449972488775,
            "support": 547.0
          },
          "weighted avg": {
            "precision": 0.3887755169626113,
            "recall": 0.376599634369287,
            "f1-score": 0.3728544463898692,
            "support": 547.0
          }
        },
        "prediction_counts": [
          271,
          153,
          123
        ]
      },
      "swap_average": {
        "rows": 547,
        "log_loss": 1.0823390483856201,
        "accuracy": 0.4149908592321755,
        "macro_f1": 0.41049598022277883,
        "brier_sum_three_classes": 0.6549387023135711,
        "per_class": {
          "a_wins": {
            "precision": 0.42328042328042326,
            "recall": 0.43243243243243246,
            "f1-score": 0.42780748663101603,
            "support": 185.0
          },
          "b_wins": {
            "precision": 0.4842105263157895,
            "recall": 0.43601895734597157,
            "f1-score": 0.45885286783042395,
            "support": 211.0
          },
          "tie": {
            "precision": 0.3273809523809524,
            "recall": 0.36423841059602646,
            "f1-score": 0.3448275862068966,
            "support": 151.0
          },
          "accuracy": 0.4149908592321755,
          "macro avg": {
            "precision": 0.41162396732572165,
            "recall": 0.41089660012481016,
            "f1-score": 0.41049598022277883,
            "support": 547.0
          },
          "weighted avg": {
            "precision": 0.42031046283187146,
            "recall": 0.4149908592321755,
            "f1-score": 0.41687624434405635,
            "support": 547.0
          }
        },
        "prediction_counts": [
          189,
          190,
          168
        ]
      }
    },
    "untruncated": {
      "uniform": {
        "rows": 453,
        "log_loss": 1.0986122886681096,
        "accuracy": 0.3443708609271523,
        "macro_f1": 0.17077175697865354,
        "brier_sum_three_classes": 0.6666666666666666,
        "per_class": {
          "a_wins": {
            "precision": 0.3443708609271523,
            "recall": 1.0,
            "f1-score": 0.5123152709359606,
            "support": 156.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 153.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 144.0
          },
          "accuracy": 0.3443708609271523,
          "macro avg": {
            "precision": 0.11479028697571743,
            "recall": 0.3333333333333333,
            "f1-score": 0.17077175697865354,
            "support": 453.0
          },
          "weighted avg": {
            "precision": 0.11859128985570806,
            "recall": 0.3443708609271523,
            "f1-score": 0.17642645091834408,
            "support": 453.0
          }
        },
        "prediction_counts": [
          453,
          0,
          0
        ]
      },
      "train_prior": {
        "rows": 453,
        "log_loss": 1.0984476392221993,
        "accuracy": 0.33774834437086093,
        "macro_f1": 0.16831683168316833,
        "brier_sum_three_classes": 0.6665541434188742,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 156.0
          },
          "b_wins": {
            "precision": 0.33774834437086093,
            "recall": 1.0,
            "f1-score": 0.504950495049505,
            "support": 153.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 144.0
          },
          "accuracy": 0.33774834437086093,
          "macro avg": {
            "precision": 0.11258278145695365,
            "recall": 0.3333333333333333,
            "f1-score": 0.16831683168316833,
            "support": 453.0
          },
          "weighted avg": {
            "precision": 0.11407394412525766,
            "recall": 0.33774834437086093,
            "f1-score": 0.1705461936922169,
            "support": 453.0
          }
        },
        "prediction_counts": [
          0,
          453,
          0
        ]
      },
      "single": {
        "rows": 453,
        "log_loss": 1.0687402486801147,
        "accuracy": 0.4326710816777042,
        "macro_f1": 0.43029545151393894,
        "brier_sum_three_classes": 0.6444663926602061,
        "per_class": {
          "a_wins": {
            "precision": 0.4492753623188406,
            "recall": 0.3974358974358974,
            "f1-score": 0.4217687074829932,
            "support": 156.0
          },
          "b_wins": {
            "precision": 0.46218487394957986,
            "recall": 0.35947712418300654,
            "f1-score": 0.40441176470588236,
            "support": 153.0
          },
          "tie": {
            "precision": 0.4030612244897959,
            "recall": 0.5486111111111112,
            "f1-score": 0.4647058823529412,
            "support": 144.0
          },
          "accuracy": 0.4326710816777042,
          "macro avg": {
            "precision": 0.43817382025273877,
            "recall": 0.43517471091000504,
            "f1-score": 0.43029545151393894,
            "support": 453.0
          },
          "weighted avg": {
            "precision": 0.43894494163919523,
            "recall": 0.4326710816777042,
            "f1-score": 0.4295553320666015,
            "support": 453.0
          }
        },
        "prediction_counts": [
          138,
          119,
          196
        ]
      },
      "swap_average": {
        "rows": 453,
        "log_loss": 1.0585500001907349,
        "accuracy": 0.445916114790287,
        "macro_f1": 0.4440552337769759,
        "brier_sum_three_classes": 0.6379344644281317,
        "per_class": {
          "a_wins": {
            "precision": 0.5,
            "recall": 0.391025641025641,
            "f1-score": 0.43884892086330934,
            "support": 156.0
          },
          "b_wins": {
            "precision": 0.49572649572649574,
            "recall": 0.3790849673202614,
            "f1-score": 0.42962962962962964,
            "support": 153.0
          },
          "tie": {
            "precision": 0.3878504672897196,
            "recall": 0.5763888888888888,
            "f1-score": 0.46368715083798884,
            "support": 144.0
          },
          "accuracy": 0.445916114790287,
          "macro avg": {
            "precision": 0.46119232100540514,
            "recall": 0.44883316574493043,
            "f1-score": 0.4440552337769759,
            "support": 453.0
          },
          "weighted avg": {
            "precision": 0.4629064484235618,
            "recall": 0.445916114790287,
            "f1-score": 0.4436307167962031,
            "support": 453.0
          }
        },
        "prediction_counts": [
          122,
          117,
          214
        ]
      }
    },
    "multi_turn": {
      "uniform": {
        "rows": 115,
        "log_loss": 1.0986122886681093,
        "accuracy": 0.33043478260869563,
        "macro_f1": 0.1655773420479303,
        "brier_sum_three_classes": 0.6666666666666669,
        "per_class": {
          "a_wins": {
            "precision": 0.33043478260869563,
            "recall": 1.0,
            "f1-score": 0.49673202614379086,
            "support": 38.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 47.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 30.0
          },
          "accuracy": 0.33043478260869563,
          "macro avg": {
            "precision": 0.11014492753623188,
            "recall": 0.3333333333333333,
            "f1-score": 0.1655773420479303,
            "support": 115.0
          },
          "weighted avg": {
            "precision": 0.10918714555765595,
            "recall": 0.33043478260869563,
            "f1-score": 0.16413753907360046,
            "support": 115.0
          }
        },
        "prediction_counts": [
          115,
          0,
          0
        ]
      },
      "train_prior": {
        "rows": 115,
        "log_loss": 1.0907540689967332,
        "accuracy": 0.40869565217391307,
        "macro_f1": 0.1934156378600823,
        "brier_sum_three_classes": 0.6615083328804348,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 38.0
          },
          "b_wins": {
            "precision": 0.40869565217391307,
            "recall": 1.0,
            "f1-score": 0.5802469135802469,
            "support": 47.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 30.0
          },
          "accuracy": 0.40869565217391307,
          "macro avg": {
            "precision": 0.13623188405797101,
            "recall": 0.3333333333333333,
            "f1-score": 0.1934156378600823,
            "support": 115.0
          },
          "weighted avg": {
            "precision": 0.16703213610586012,
            "recall": 0.40869565217391307,
            "f1-score": 0.23714439076757918,
            "support": 115.0
          }
        },
        "prediction_counts": [
          0,
          115,
          0
        ]
      },
      "single": {
        "rows": 115,
        "log_loss": 1.0935041904449463,
        "accuracy": 0.3130434782608696,
        "macro_f1": 0.28258359268812233,
        "brier_sum_three_classes": 0.6628708324643462,
        "per_class": {
          "a_wins": {
            "precision": 0.3114754098360656,
            "recall": 0.5,
            "f1-score": 0.3838383838383838,
            "support": 38.0
          },
          "b_wins": {
            "precision": 0.4,
            "recall": 0.2978723404255319,
            "f1-score": 0.34146341463414637,
            "support": 47.0
          },
          "tie": {
            "precision": 0.15789473684210525,
            "recall": 0.1,
            "f1-score": 0.12244897959183673,
            "support": 30.0
          },
          "accuracy": 0.3130434782608696,
          "macro avg": {
            "precision": 0.28979004889272364,
            "recall": 0.29929078014184396,
            "f1-score": 0.28258359268812233,
            "support": 115.0
          },
          "weighted avg": {
            "precision": 0.3075905015568144,
            "recall": 0.3130434782608696,
            "f1-score": 0.2983313779253789,
            "support": 115.0
          }
        },
        "prediction_counts": [
          61,
          35,
          19
        ]
      },
      "swap_average": {
        "rows": 115,
        "log_loss": 1.097314715385437,
        "accuracy": 0.3826086956521739,
        "macro_f1": 0.3708399787347156,
        "brier_sum_three_classes": 0.6653494928998362,
        "per_class": {
          "a_wins": {
            "precision": 0.38,
            "recall": 0.5,
            "f1-score": 0.4318181818181818,
            "support": 38.0
          },
          "b_wins": {
            "precision": 0.4473684210526316,
            "recall": 0.3617021276595745,
            "f1-score": 0.4,
            "support": 47.0
          },
          "tie": {
            "precision": 0.2962962962962963,
            "recall": 0.26666666666666666,
            "f1-score": 0.2807017543859649,
            "support": 30.0
          },
          "accuracy": 0.3826086956521739,
          "macro avg": {
            "precision": 0.37455490578297596,
            "recall": 0.37612293144208037,
            "f1-score": 0.3708399787347156,
            "support": 115.0
          },
          "weighted avg": {
            "precision": 0.38569743198576145,
            "recall": 0.3826086956521739,
            "f1-score": 0.379392552527564,
            "support": 115.0
          }
        },
        "prediction_counts": [
          50,
          38,
          27
        ]
      }
    },
    "label_0": {
      "uniform": {
        "rows": 341,
        "log_loss": 1.0986122886681098,
        "accuracy": 1.0,
        "macro_f1": 0.3333333333333333,
        "brier_sum_three_classes": 0.6666666666666667,
        "per_class": {
          "a_wins": {
            "precision": 1.0,
            "recall": 1.0,
            "f1-score": 1.0,
            "support": 341.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "accuracy": 1.0,
          "macro avg": {
            "precision": 0.3333333333333333,
            "recall": 0.3333333333333333,
            "f1-score": 0.3333333333333333,
            "support": 341.0
          },
          "weighted avg": {
            "precision": 1.0,
            "recall": 1.0,
            "f1-score": 1.0,
            "support": 341.0
          }
        },
        "prediction_counts": [
          341,
          0,
          0
        ]
      },
      "train_prior": {
        "rows": 341,
        "log_loss": 1.0663871413733723,
        "accuracy": 0.0,
        "macro_f1": 0.0,
        "brier_sum_three_classes": 0.6459365937500002,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 341.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "accuracy": 0.0,
          "macro avg": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 341.0
          },
          "weighted avg": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 341.0
          }
        },
        "prediction_counts": [
          0,
          341,
          0
        ]
      },
      "single": {
        "rows": 341,
        "log_loss": 1.037455677986145,
        "accuracy": 0.45454545454545453,
        "macro_f1": 0.20833333333333334,
        "brier_sum_three_classes": 0.626283933193414,
        "per_class": {
          "a_wins": {
            "precision": 1.0,
            "recall": 0.45454545454545453,
            "f1-score": 0.625,
            "support": 341.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "accuracy": 0.45454545454545453,
          "macro avg": {
            "precision": 0.3333333333333333,
            "recall": 0.15151515151515152,
            "f1-score": 0.20833333333333334,
            "support": 341.0
          },
          "weighted avg": {
            "precision": 1.0,
            "recall": 0.45454545454545453,
            "f1-score": 0.625,
            "support": 341.0
          }
        },
        "prediction_counts": [
          155,
          86,
          100
        ]
      },
      "swap_average": {
        "rows": 341,
        "log_loss": 1.0757343769073486,
        "accuracy": 0.41348973607038125,
        "macro_f1": 0.1950207468879668,
        "brier_sum_three_classes": 0.6515761313666348,
        "per_class": {
          "a_wins": {
            "precision": 1.0,
            "recall": 0.41348973607038125,
            "f1-score": 0.5850622406639004,
            "support": 341.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "accuracy": 0.41348973607038125,
          "macro avg": {
            "precision": 0.3333333333333333,
            "recall": 0.13782991202346043,
            "f1-score": 0.1950207468879668,
            "support": 341.0
          },
          "weighted avg": {
            "precision": 1.0,
            "recall": 0.41348973607038125,
            "f1-score": 0.5850622406639004,
            "support": 341.0
          }
        },
        "prediction_counts": [
          141,
          85,
          115
        ]
      }
    },
    "label_1": {
      "uniform": {
        "rows": 364,
        "log_loss": 1.0986122886681096,
        "accuracy": 0.0,
        "macro_f1": 0.0,
        "brier_sum_three_classes": 0.6666666666666669,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 364.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "accuracy": 0.0,
          "macro avg": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 364.0
          },
          "weighted avg": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 364.0
          }
        },
        "prediction_counts": [
          364,
          0,
          0
        ]
      },
      "train_prior": {
        "rows": 364,
        "log_loss": 1.0516094350727736,
        "accuracy": 1.0,
        "macro_f1": 0.3333333333333333,
        "brier_sum_three_classes": 0.63568659375,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "b_wins": {
            "precision": 1.0,
            "recall": 1.0,
            "f1-score": 1.0,
            "support": 364.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "accuracy": 1.0,
          "macro avg": {
            "precision": 0.3333333333333333,
            "recall": 0.3333333333333333,
            "f1-score": 0.3333333333333333,
            "support": 364.0
          },
          "weighted avg": {
            "precision": 1.0,
            "recall": 1.0,
            "f1-score": 1.0,
            "support": 364.0
          }
        },
        "prediction_counts": [
          0,
          364,
          0
        ]
      },
      "single": {
        "rows": 364,
        "log_loss": 1.1268516778945923,
        "accuracy": 0.34065934065934067,
        "macro_f1": 0.16939890710382513,
        "brier_sum_three_classes": 0.6820530016857571,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "b_wins": {
            "precision": 1.0,
            "recall": 0.34065934065934067,
            "f1-score": 0.5081967213114754,
            "support": 364.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "accuracy": 0.34065934065934067,
          "macro avg": {
            "precision": 0.3333333333333333,
            "recall": 0.11355311355311355,
            "f1-score": 0.16939890710382513,
            "support": 364.0
          },
          "weighted avg": {
            "precision": 1.0,
            "recall": 0.34065934065934067,
            "f1-score": 0.5081967213114754,
            "support": 364.0
          }
        },
        "prediction_counts": [
          144,
          124,
          96
        ]
      },
      "swap_average": {
        "rows": 364,
        "log_loss": 1.066325068473816,
        "accuracy": 0.41208791208791207,
        "macro_f1": 0.19455252918287938,
        "brier_sum_three_classes": 0.645313366892336,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "b_wins": {
            "precision": 1.0,
            "recall": 0.41208791208791207,
            "f1-score": 0.5836575875486382,
            "support": 364.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "accuracy": 0.41208791208791207,
          "macro avg": {
            "precision": 0.3333333333333333,
            "recall": 0.13736263736263735,
            "f1-score": 0.19455252918287938,
            "support": 364.0
          },
          "weighted avg": {
            "precision": 1.0,
            "recall": 0.41208791208791207,
            "f1-score": 0.5836575875486382,
            "support": 364.0
          }
        },
        "prediction_counts": [
          85,
          150,
          129
        ]
      }
    },
    "label_2": {
      "uniform": {
        "rows": 295,
        "log_loss": 1.0986122886681096,
        "accuracy": 0.0,
        "macro_f1": 0.0,
        "brier_sum_three_classes": 0.6666666666666669,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 295.0
          },
          "accuracy": 0.0,
          "macro avg": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 295.0
          },
          "weighted avg": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 295.0
          }
        },
        "prediction_counts": [
          295,
          0,
          0
        ]
      },
      "train_prior": {
        "rows": 295,
        "log_loss": 1.1829454371338604,
        "accuracy": 0.0,
        "macro_f1": 0.0,
        "brier_sum_three_classes": 0.7216865937499999,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 295.0
          },
          "accuracy": 0.0,
          "macro avg": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 295.0
          },
          "weighted avg": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 295.0
          }
        },
        "prediction_counts": [
          0,
          295,
          0
        ]
      },
      "single": {
        "rows": 295,
        "log_loss": 1.0724611282348633,
        "accuracy": 0.41694915254237286,
        "macro_f1": 0.19617224880382775,
        "brier_sum_three_classes": 0.6479630737293993,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "tie": {
            "precision": 1.0,
            "recall": 0.41694915254237286,
            "f1-score": 0.5885167464114832,
            "support": 295.0
          },
          "accuracy": 0.41694915254237286,
          "macro avg": {
            "precision": 0.3333333333333333,
            "recall": 0.13898305084745763,
            "f1-score": 0.19617224880382775,
            "support": 295.0
          },
          "weighted avg": {
            "precision": 1.0,
            "recall": 0.41694915254237286,
            "f1-score": 0.5885167464114832,
            "support": 295.0
          }
        },
        "prediction_counts": [
          110,
          62,
          123
        ]
      },
      "swap_average": {
        "rows": 295,
        "log_loss": 1.0732029676437378,
        "accuracy": 0.46779661016949153,
        "macro_f1": 0.21247113163972287,
        "brier_sum_three_classes": 0.6445906990055401,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 0.0
          },
          "tie": {
            "precision": 1.0,
            "recall": 0.46779661016949153,
            "f1-score": 0.6374133949191686,
            "support": 295.0
          },
          "accuracy": 0.46779661016949153,
          "macro avg": {
            "precision": 0.3333333333333333,
            "recall": 0.15593220338983052,
            "f1-score": 0.21247113163972287,
            "support": 295.0
          },
          "weighted avg": {
            "precision": 1.0,
            "recall": 0.46779661016949153,
            "f1-score": 0.6374133949191686,
            "support": 295.0
          }
        },
        "prediction_counts": [
          85,
          72,
          138
        ]
      }
    },
    "packed_under256": {
      "uniform": {
        "rows": 210,
        "log_loss": 1.0986122886681096,
        "accuracy": 0.3142857142857143,
        "macro_f1": 0.15942028985507248,
        "brier_sum_three_classes": 0.6666666666666669,
        "per_class": {
          "a_wins": {
            "precision": 0.3142857142857143,
            "recall": 1.0,
            "f1-score": 0.4782608695652174,
            "support": 66.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 69.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 75.0
          },
          "accuracy": 0.3142857142857143,
          "macro avg": {
            "precision": 0.10476190476190476,
            "recall": 0.3333333333333333,
            "f1-score": 0.15942028985507248,
            "support": 210.0
          },
          "weighted avg": {
            "precision": 0.09877551020408164,
            "recall": 0.3142857142857143,
            "f1-score": 0.15031055900621118,
            "support": 210.0
          }
        },
        "prediction_counts": [
          210,
          0,
          0
        ]
      },
      "train_prior": {
        "rows": 210,
        "log_loss": 1.1031595720747784,
        "accuracy": 0.32857142857142857,
        "macro_f1": 0.16487455197132617,
        "brier_sum_three_classes": 0.6696223080357142,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 66.0
          },
          "b_wins": {
            "precision": 0.32857142857142857,
            "recall": 1.0,
            "f1-score": 0.4946236559139785,
            "support": 69.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 75.0
          },
          "accuracy": 0.32857142857142857,
          "macro avg": {
            "precision": 0.10952380952380952,
            "recall": 0.3333333333333333,
            "f1-score": 0.16487455197132617,
            "support": 210.0
          },
          "weighted avg": {
            "precision": 0.10795918367346939,
            "recall": 0.32857142857142857,
            "f1-score": 0.16251920122887867,
            "support": 210.0
          }
        },
        "prediction_counts": [
          0,
          210,
          0
        ]
      },
      "single": {
        "rows": 210,
        "log_loss": 1.07950758934021,
        "accuracy": 0.38571428571428573,
        "macro_f1": 0.3325017937968474,
        "brier_sum_three_classes": 0.6525853612200789,
        "per_class": {
          "a_wins": {
            "precision": 0.3793103448275862,
            "recall": 0.16666666666666666,
            "f1-score": 0.23157894736842105,
            "support": 66.0
          },
          "b_wins": {
            "precision": 0.34210526315789475,
            "recall": 0.18840579710144928,
            "f1-score": 0.24299065420560748,
            "support": 69.0
          },
          "tie": {
            "precision": 0.3986013986013986,
            "recall": 0.76,
            "f1-score": 0.5229357798165137,
            "support": 75.0
          },
          "accuracy": 0.38571428571428573,
          "macro avg": {
            "precision": 0.3733390021956265,
            "recall": 0.37169082125603864,
            "f1-score": 0.3325017937968474,
            "support": 210.0
          },
          "weighted avg": {
            "precision": 0.3739754800553348,
            "recall": 0.38571428571428573,
            "f1-score": 0.3393845197749583,
            "support": 210.0
          }
        },
        "prediction_counts": [
          29,
          38,
          143
        ]
      },
      "swap_average": {
        "rows": 210,
        "log_loss": 1.0678188800811768,
        "accuracy": 0.40476190476190477,
        "macro_f1": 0.3455772113943028,
        "brier_sum_three_classes": 0.644879600077552,
        "per_class": {
          "a_wins": {
            "precision": 0.46153846153846156,
            "recall": 0.18181818181818182,
            "f1-score": 0.2608695652173913,
            "support": 66.0
          },
          "b_wins": {
            "precision": 0.4444444444444444,
            "recall": 0.17391304347826086,
            "f1-score": 0.25,
            "support": 69.0
          },
          "tie": {
            "precision": 0.3885350318471338,
            "recall": 0.8133333333333334,
            "f1-score": 0.5258620689655172,
            "support": 75.0
          },
          "accuracy": 0.40476190476190477,
          "macro avg": {
            "precision": 0.43150597927667994,
            "recall": 0.38968818620992535,
            "f1-score": 0.3455772113943028,
            "support": 210.0
          },
          "weighted avg": {
            "precision": 0.42984920246066743,
            "recall": 0.40476190476190477,
            "f1-score": 0.3519383165560077,
            "support": 210.0
          }
        },
        "prediction_counts": [
          26,
          27,
          157
        ]
      }
    },
    "packed_256plus": {
      "uniform": {
        "rows": 790,
        "log_loss": 1.0986122886681096,
        "accuracy": 0.34810126582278483,
        "macro_f1": 0.17214397496087638,
        "brier_sum_three_classes": 0.6666666666666667,
        "per_class": {
          "a_wins": {
            "precision": 0.34810126582278483,
            "recall": 1.0,
            "f1-score": 0.5164319248826291,
            "support": 275.0
          },
          "b_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 295.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 220.0
          },
          "accuracy": 0.34810126582278483,
          "macro avg": {
            "precision": 0.11603375527426162,
            "recall": 0.3333333333333333,
            "f1-score": 0.17214397496087638,
            "support": 790.0
          },
          "weighted avg": {
            "precision": 0.1211744912674251,
            "recall": 0.34810126582278483,
            "f1-score": 0.17977060676294052,
            "support": 790.0
          }
        },
        "prediction_counts": [
          790,
          0,
          0
        ]
      },
      "train_prior": {
        "rows": 790,
        "log_loss": 1.0933281561944237,
        "accuracy": 0.37341772151898733,
        "macro_f1": 0.18125960061443933,
        "brier_sum_three_classes": 0.6632039988132912,
        "per_class": {
          "a_wins": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 275.0
          },
          "b_wins": {
            "precision": 0.37341772151898733,
            "recall": 1.0,
            "f1-score": 0.543778801843318,
            "support": 295.0
          },
          "tie": {
            "precision": 0.0,
            "recall": 0.0,
            "f1-score": 0.0,
            "support": 220.0
          },
          "accuracy": 0.37341772151898733,
          "macro avg": {
            "precision": 0.12447257383966244,
            "recall": 0.3333333333333333,
            "f1-score": 0.18125960061443933,
            "support": 790.0
          },
          "weighted avg": {
            "precision": 0.13944079474443197,
            "recall": 0.37341772151898733,
            "f1-score": 0.2030566411946567,
            "support": 790.0
          }
        },
        "prediction_counts": [
          0,
          790,
          0
        ]
      },
      "single": {
        "rows": 790,
        "log_loss": 1.0805391073226929,
        "accuracy": 0.40632911392405063,
        "macro_f1": 0.3975625750571998,
        "brier_sum_three_classes": 0.6530839173753493,
        "per_class": {
          "a_wins": {
            "precision": 0.37894736842105264,
            "recall": 0.5236363636363637,
            "f1-score": 0.4396946564885496,
            "support": 275.0
          },
          "b_wins": {
            "precision": 0.47435897435897434,
            "recall": 0.376271186440678,
            "f1-score": 0.41965973534971646,
            "support": 295.0
          },
          "tie": {
            "precision": 0.375,
            "recall": 0.3,
            "f1-score": 0.3333333333333333,
            "support": 220.0
          },
          "accuracy": 0.40632911392405063,
          "macro avg": {
            "precision": 0.4094354475933423,
            "recall": 0.39996918335901394,
            "f1-score": 0.3975625750571998,
            "support": 790.0
          },
          "weighted avg": {
            "precision": 0.41347648576162904,
            "recall": 0.40632911392405063,
            "f1-score": 0.40259365290614024,
            "support": 790.0
          }
        },
        "prediction_counts": [
          380,
          234,
          176
        ]
      },
      "swap_average": {
        "rows": 790,
        "log_loss": 1.072557806968689,
        "accuracy": 0.43544303797468353,
        "macro_f1": 0.42892723381487424,
        "brier_sum_three_classes": 0.6478621095382041,
        "per_class": {
          "a_wins": {
            "precision": 0.45263157894736844,
            "recall": 0.4690909090909091,
            "f1-score": 0.4607142857142857,
            "support": 275.0
          },
          "b_wins": {
            "precision": 0.4928571428571429,
            "recall": 0.46779661016949153,
            "f1-score": 0.48,
            "support": 295.0
          },
          "tie": {
            "precision": 0.3422222222222222,
            "recall": 0.35,
            "f1-score": 0.34606741573033706,
            "support": 220.0
          },
          "accuracy": 0.43544303797468353,
          "macro avg": {
            "precision": 0.42923698134224453,
            "recall": 0.4289625064201335,
            "f1-score": 0.42892723381487424,
            "support": 790.0
          },
          "weighted avg": {
            "precision": 0.4369056079016106,
            "recall": 0.43544303797468353,
            "f1-score": 0.4359889367494971,
            "support": 790.0
          }
        },
        "prediction_counts": [
          285,
          280,
          225
        ]
      }
    }
  },
  "failures": [
    {
      "id": "1811870726",
      "human_label": 2,
      "predicted": "a_wins",
      "probabilities": [
        0.5536031126976013,
        0.28300854563713074,
        0.16338837146759033
      ],
      "truncated": false,
      "note": "high-loss mismatch; retrieve authorized source by ID to inspect, no conversation text uploaded"
    },
    {
      "id": "3774524045",
      "human_label": 2,
      "predicted": "a_wins",
      "probabilities": [
        0.5416786670684814,
        0.2832186222076416,
        0.17510277032852173
      ],
      "truncated": false,
      "note": "high-loss mismatch; retrieve authorized source by ID to inspect, no conversation text uploaded"
    },
    {
      "id": "2936859080",
      "human_label": 2,
      "predicted": "a_wins",
      "probabilities": [
        0.4162439703941345,
        0.39672791957855225,
        0.18702811002731323
      ],
      "truncated": false,
      "note": "high-loss mismatch; retrieve authorized source by ID to inspect, no conversation text uploaded"
    },
    {
      "id": "3682941009",
      "human_label": 2,
      "predicted": "b_wins",
      "probabilities": [
        0.37328940629959106,
        0.438693106174469,
        0.18801751732826233
      ],
      "truncated": false,
      "note": "high-loss mismatch; retrieve authorized source by ID to inspect, no conversation text uploaded"
    },
    {
      "id": "2122284002",
      "human_label": 1,
      "predicted": "a_wins",
      "probabilities": [
        0.5208093523979187,
        0.1932239830493927,
        0.285966694355011
      ],
      "truncated": false,
      "note": "high-loss mismatch; retrieve authorized source by ID to inspect, no conversation text uploaded"
    }
  ],
  "cost": {
    "single": {
      "one_pair": {
        "repetitions": 3,
        "rows": 1,
        "seconds": [
          0.04551020008511841,
          0.048418299993500113,
          0.04612740012817085
        ],
        "median_seconds": 0.04612740012817085,
        "pairs_per_second": 21.6790887243021,
        "peak_cuda_allocated_bytes": [
          2011312640,
          2011312640,
          2011312640
        ],
        "includes": "CPU packing/padding and GPU forward, synchronized; excludes loading and warmup"
      },
      "batch64": {
        "repetitions": 3,
        "rows": 64,
        "seconds": [
          2.3503066000994295,
          2.347085600020364,
          2.3406206001527607
        ],
        "median_seconds": 2.347085600020364,
        "pairs_per_second": 27.267859339874406,
        "peak_cuda_allocated_bytes": [
          2172357632,
          2172357632,
          2172357632
        ],
        "includes": "CPU packing/padding and GPU forward, synchronized; excludes loading and warmup"
      },
      "gpu_forward_intervals_batch64": {
        "measured": true,
        "milliseconds": [
          2506.2051544189453,
          2505.1741333007812,
          2508.885025024414
        ],
        "includes": "GPU-event forward intervals after padding/device transfer, three repetitions; shared GPU contention may affect intervals"
      }
    },
    "swap_average": {
      "one_pair": {
        "repetitions": 3,
        "rows": 1,
        "seconds": [
          0.09537050011567771,
          0.10050359996967018,
          0.08934069983661175
        ],
        "median_seconds": 0.09537050011567771,
        "pairs_per_second": 10.485422628455028,
        "peak_cuda_allocated_bytes": [
          2011312640,
          2011312640,
          2011312640
        ],
        "includes": "CPU packing/padding and GPU forward, synchronized; excludes loading and warmup"
      },
      "batch64": {
        "repetitions": 3,
        "rows": 64,
        "seconds": [
          4.7567368999589235,
          4.795195200014859,
          4.750788100063801
        ],
        "median_seconds": 4.7567368999589235,
        "pairs_per_second": 13.454601619978744,
        "peak_cuda_allocated_bytes": [
          2172357632,
          2172357632,
          2172357632
        ],
        "includes": "CPU packing/padding and GPU forward, synchronized; excludes loading and warmup"
      },
      "gpu_forward_intervals_batch64": {
        "measured": true,
        "milliseconds": [
          5329.472396850586,
          5014.668655395508,
          5010.498687744141
        ],
        "includes": "GPU-event forward intervals after padding/device transfer, three repetitions; shared GPU contention may affect intervals"
      }
    }
  },
  "loading_seconds": 5.098238100064918,
  "two_test_passes_seconds": 75.38759940001182,
  "hardware": {
    "gpu": "NVIDIA GeForce RTX 4080",
    "os": "Windows-11-10.0.22631-SP0",
    "torch": "2.6.0+cu124",
    "device": "cuda:0",
    "dtype": "float32",
    "batch_size": 4
  },
  "training": {
    "optimizer_steps": 500,
    "head_changed": true,
    "head_before": {
      "base_model.model.score.modules_to_save.default.weight": "5e13678b15acea2a29dd4754bd62b5ed9e3def1e5579899eec76a6700c2dd587"
    },
    "head_after": {
      "base_model.model.score.modules_to_save.default.weight": "a3f7321c4eb3ef605d46d7473e473232378b45d7fbef70bf22bbf5ad1aa9864c"
    },
    "seed": 42,
    "label_mode": "hard",
    "train_rows": 8000,
    "validation_rows": 769,
    "history": [
      {
        "step": 500,
        "log_loss": 1.0999491249243831,
        "acc": 0.3784135240572172
      }
    ],
    "elapsed_s": 1230.666043000063,
    "device": "cuda:0",
    "dtype": "torch.bfloat16",
    "peak_cuda_allocated_bytes": 1373425152,
    "config": {
      "label_mode": "hard",
      "eval_holdout": 0.05,
      "seed": 42,
      "model_revision": "7ae557604adf67be50417f59c2c2f167def9a775",
      "local_files_only": true,
      "max_length": 512,
      "packer": {},
      "label_order": [
        "a_wins",
        "b_wins",
        "tie"
      ],
      "disable_attn_logit_softcapping": false,
      "use_lora": true,
      "lora_r": 16,
      "lora_alpha": 32,
      "lora_dropout": 0.05,
      "lora_target_modules": [
        "q_proj",
        "k_proj",
        "v_proj",
        "o_proj",
        "gate_proj",
        "up_proj",
        "down_proj"
      ],
      "n_epochs": 1.0,
      "max_steps": 0,
      "lr": 0.0001,
      "per_device_train_batch_size": 2,
      "per_device_eval_batch_size": 4,
      "gradient_accumulation_steps": 8,
      "gradient_checkpointing": true,
      "warmup_steps": 20,
      "optim": "adamw_torch",
      "device": "cuda:0",
      "dtype": "bfloat16",
      "bf16": false,
      "num_workers": 0,
      "merge_adapter": true,
      "provenance": {
        "dataset_id": "lmarena-ai/arena-human-preference-55k",
        "dataset_license": "apache-2.0",
        "dataset_revision": "18c298340948c0e7f7727399fd459cca6ce0ca6f",
        "license": "apache-2.0",
        "purpose": "public example trained on frozen human labels",
        "split_manifest_sha256": "28106a452c911228f89bfdbf802edfafab75e45f417f269ac65bc51d4dbf438d"
      }
    },
    "train_file_sha256": "e92c74996c7d55beb1faaeafc6936bf0a3b2313026b981d564c2a7bdf680f9ae",
    "validation_file_sha256": "bfd7bb542243c5a76ed830d0ab3fc82261cb206e239e046c2fac4c8b15550236"
  },
  "artifact_files": {
    "added_tokens.json": "f0acedf99e19ec0a5b797ffbd0b94328172ccd90f7f385bbc521cda2701b03ab",
    "chat_template.jinja": "8aa40ce145adb73cb3a75194dc0224702a95850ec5275cabb728496bbd749fc6",
    "config.json": "8e0e4ace7723832aaf5968b91ed6375b797dcbf2666e7eb1c3f0d8ae1542262a",
    "merges.txt": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5",
    "model.safetensors": "0c2c3b33574d8e761b0b3ff088268068cfef20333943e4e1cac6c55edcf767a9",
    "special_tokens_map.json": "57255613bbe23c9497211ca68561ff429a51e871dbaf5a59998fa4c8f7fe168a",
    "tokenizer.json": "9c5ae00e602b8860cbd784ba82a8aa14e8feecec692e7076590d014d7b7fdafa",
    "tokenizer_config.json": "f487c36ca6195f07ca7d72a2836d2575840124f400ed7c4c3d5fd3015d434790",
    "vocab.json": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
  },
  "split_manifest_sha256": "28106a452c911228f89bfdbf802edfafab75e45f417f269ac65bc51d4dbf438d",
  "prediction_file_sha256": "7f502ff4c7f1ded200b5c94446422d079223254c74751ac5c4fa379227c2e1ff",
  "quality_gate": {
    "beats_uniform_and_train_prior_log_loss": true,
    "nonconstant_predictions": true,
    "finite_three_class_output": true
  },
  "limitations": "Uncalibrated; one dataset; no external contamination guarantee. Bootstrap measures test-sampling uncertainty, not training variability. Costs measured while sharing GPU with existing work."
}