{
  "batch": 8,
  "direct_hit3_threshold": 0.9,
  "direct_report": {
    "final_dir": "/content/yandex_disk_cache/ml/subgroup_distillation/group_e/direct/alexyalunin-rubiobert/final",
    "model_name": "alexyalunin/RuBioBERT",
    "run_slug": "direct_hard_training",
    "selection_metric": "hit@3",
    "selection_metric_value": 0.835,
    "test_metrics": {
      "hit@1": 0.628140703517588,
      "hit@3": 0.8291457286432161,
      "macro_f1": 0.5157875868327818,
      "micro_f1": 0.48293963254593175,
      "mrr": 0.744193071244231,
      "recall@3": 0.8197654941373534,
      "subset_accuracy": 0.1708542713567839,
      "weighted_f1": 0.5126833520048437
    },
    "train_duration_sec": 100.4,
    "train_metrics": {
      "epoch": 9.0,
      "total_flos": 2177031722148864.0,
      "train_loss": 0.47315873063128927,
      "train_runtime": 100.0873,
      "train_samples_per_second": 110.184,
      "train_steps_per_second": 13.788
    },
    "val_metrics": {
      "hit@1": 0.645,
      "hit@3": 0.835,
      "macro_f1": 0.5239739935027187,
      "micro_f1": 0.48812664907651715,
      "mrr": 0.750362070235822,
      "recall@3": 0.8273333333333333,
      "subset_accuracy": 0.175,
      "weighted_f1": 0.5181260346642617
    }
  },
  "dropout": 0.1,
  "epochs": 12,
  "generated_at_utc": "2026-05-26T08:05:56+00:00",
  "group_letter": "E",
  "kd_grid": [
    {
      "hard_loss_weight": 0.5,
      "temperature": 2.0
    },
    {
      "hard_loss_weight": 0.3,
      "temperature": 2.0
    },
    {
      "hard_loss_weight": 0.3,
      "temperature": 2.5
    }
  ],
  "lr": 2e-05,
  "max_len": 512,
  "pos_weight_cap": 20.0,
  "pos_weight_stats": {
    "labels_at_cap": 39,
    "labels_without_positives": 0,
    "max": 20.0,
    "mean": 18.944591522216797,
    "min": 4.536144733428955
  },
  "seed": 42,
  "selected_run": {
    "final_dir": "/content/yandex_disk_cache/ml/subgroup_distillation/group_e/runs/alexyalunin-rubiobert/t2p0_hard0p5/final",
    "hard_loss_weight": 0.5,
    "run_slug": "t2p0_hard0p5",
    "selection_metric": "val_macro_f1",
    "selection_metric_value": 0.6257847034784294,
    "temperature": 2.0,
    "test_metrics": {
      "test_hit@1": 0.6482412060301508,
      "test_hit@3": 0.8241206030150754,
      "test_loss": 1.2573049068450928,
      "test_macro_f1": 0.6121827915051391,
      "test_micro_f1": 0.5606299212598426,
      "test_mrr": 0.7515661740428731,
      "test_recall@3": 0.8147403685092127,
      "test_runtime": 0.4221,
      "test_samples_per_second": 471.485,
      "test_steps_per_second": 30.801,
      "test_subset_accuracy": 0.2964824120603015,
      "test_weighted_f1": 0.5830428669342762
    },
    "train_duration_sec": 133.0,
    "train_metrics": {
      "epoch": 12.0,
      "total_flos": 2902708962865152.0,
      "train_loss": 1.3432866200156834,
      "train_runtime": 132.7733,
      "train_samples_per_second": 83.059,
      "train_steps_per_second": 10.394
    },
    "val_metrics": {
      "epoch": 12.0,
      "val_hit@1": 0.635,
      "val_hit@3": 0.8,
      "val_loss": 1.2582203149795532,
      "val_macro_f1": 0.6257847034784294,
      "val_micro_f1": 0.5578446909667195,
      "val_mrr": 0.7417874066293184,
      "val_recall@3": 0.7939999999999999,
      "val_runtime": 0.458,
      "val_samples_per_second": 436.701,
      "val_steps_per_second": 28.386,
      "val_subset_accuracy": 0.3,
      "val_weighted_f1": 0.581422089009702
    }
  },
  "source_csv": "/content/yandex_disk_cache/datasets/subgroups/group_E.csv",
  "source_csv_sha256": "7bd98fc0eea937b8edf1391e86ca15afd2aed5c98996951f822684805713ed0b",
  "specialist_repo_id": "Dmitry43243242/icd10-ru-subgroup-e",
  "split_sizes": {
    "test": 199,
    "train": 919,
    "val": 200
  },
  "student_model": "alexyalunin/RuBioBERT",
  "teacher_errors": {},
  "teacher_models": [
    "alexyalunin/RuBioRoBERTa",
    "ai-forever/ruBert-base",
    "DeepPavlov/rubert-base-cased"
  ],
  "teacher_reports": [
    {
      "final_dir": "/content/yandex_disk_cache/ml/subgroup_distillation/group_e/teachers/alexyalunin-rubioroberta/final",
      "model_name": "alexyalunin/RuBioRoBERTa",
      "pushed_to_hub": false,
      "run_slug": "alexyalunin-rubioroberta",
      "test_metrics": {
        "hit@1": 0.6482412060301508,
        "hit@3": 0.8592964824120602,
        "macro_f1": 0.6865864313407604,
        "micro_f1": 0.6485436893203883,
        "mrr": 0.7635155763702912,
        "recall@3": 0.8532663316582914,
        "subset_accuracy": 0.507537688442211,
        "weighted_f1": 0.6514279861261107
      },
      "train_duration_sec": 314.8,
      "train_metrics": {
        "epoch": 12.0,
        "total_flos": 1.0278832248336384e+16,
        "train_loss": 0.31829168511473616,
        "train_runtime": 299.4665,
        "train_samples_per_second": 36.825,
        "train_steps_per_second": 4.608
      },
      "val_metrics": {
        "hit@1": 0.665,
        "hit@3": 0.83,
        "macro_f1": 0.6816071208348493,
        "micro_f1": 0.6377358490566037,
        "mrr": 0.7651676992824679,
        "recall@3": 0.8214999999999999,
        "subset_accuracy": 0.48,
        "weighted_f1": 0.654032532287305
      }
    },
    {
      "final_dir": "/content/yandex_disk_cache/ml/subgroup_distillation/group_e/teachers/ai-forever-rubert-base/final",
      "model_name": "ai-forever/ruBert-base",
      "pushed_to_hub": false,
      "run_slug": "ai-forever-rubert-base",
      "test_metrics": {
        "hit@1": 0.6633165829145728,
        "hit@3": 0.8241206030150754,
        "macro_f1": 0.567726321569097,
        "micro_f1": 0.5497630331753555,
        "mrr": 0.7552527410980668,
        "recall@3": 0.8164154103852594,
        "subset_accuracy": 0.3015075376884422,
        "weighted_f1": 0.5636139940062055
      },
      "train_duration_sec": 144.6,
      "train_metrics": {
        "epoch": 12.0,
        "total_flos": 2902708962865152.0,
        "train_loss": 0.4238858779271444,
        "train_runtime": 133.7006,
        "train_samples_per_second": 82.483,
        "train_steps_per_second": 10.322
      },
      "val_metrics": {
        "hit@1": 0.65,
        "hit@3": 0.82,
        "macro_f1": 0.5512155779574014,
        "micro_f1": 0.5529953917050692,
        "mrr": 0.7548125116338592,
        "recall@3": 0.814,
        "subset_accuracy": 0.315,
        "weighted_f1": 0.5726346327425935
      }
    },
    {
      "final_dir": "/content/yandex_disk_cache/ml/subgroup_distillation/group_e/teachers/deeppavlov-rubert-base-cased/final",
      "model_name": "DeepPavlov/rubert-base-cased",
      "pushed_to_hub": false,
      "run_slug": "deeppavlov-rubert-base-cased",
      "test_metrics": {
        "hit@1": 0.6130653266331658,
        "hit@3": 0.7939698492462312,
        "macro_f1": 0.5623096108520506,
        "micro_f1": 0.5097451274362819,
        "mrr": 0.7264908372260924,
        "recall@3": 0.7879396984924623,
        "subset_accuracy": 0.2613065326633166,
        "weighted_f1": 0.5247350484683884
      },
      "train_duration_sec": 145.8,
      "train_metrics": {
        "epoch": 12.0,
        "total_flos": 2902708962865152.0,
        "train_loss": 0.45112168028734734,
        "train_runtime": 136.666,
        "train_samples_per_second": 80.693,
        "train_steps_per_second": 10.098
      },
      "val_metrics": {
        "hit@1": 0.66,
        "hit@3": 0.81,
        "macro_f1": 0.5291229985764946,
        "micro_f1": 0.5123010130246021,
        "mrr": 0.755827678749699,
        "recall@3": 0.8014999999999999,
        "subset_accuracy": 0.23,
        "weighted_f1": 0.5291843956354477
      }
    }
  ],
  "teacher_source": "local_bert_models",
  "threshold": 0.5,
  "torch_version": "2.10.0+cu128",
  "transformers_version": "5.0.0",
  "warmup_ratio": 0.1,
  "weight_decay": 0.01
}