{
  "published_checkpoint": "repository_root",
  "label_schema_version": 2,
  "label_count": 37,
  "summary": "The published checkpoint is a schema v2 character-tokenizer model. It starts from the 8-head/4-layer schema v2 baseline and finishes with a repaired hard-focus encoded-cache run that matches the repository-root model artifact hashes.",
  "summary_zh": "当前发布 checkpoint 是 37 维 schema v2 字符级模型：先训练 8-head/4-layer schema v2 baseline，再用 repaired hard-focus encoded cache 做最终 10 epoch 收口；最终产物与仓库根目录模型文件哈希一致。",
  "artifact_match": {
    "model_safetensors": "matches checkpoints/ablation-schema-v2-hardfocus-cache-repaired-from-baseline-seed62-10epoch-rerun/final/model.safetensors",
    "training_args_bin": "matches checkpoints/ablation-schema-v2-hardfocus-cache-repaired-from-baseline-seed62-10epoch-rerun/final/training_args.bin",
    "config_json": "schema v2, 37 labels, max_seq_length=128",
    "vocab_json": "char vocab, 6199 tokens"
  },
  "stages": [
    {
      "name": "ablation-schema-v2-baseline8h4l-cache-10epoch-seed62",
      "type": "schema_v2_baseline_encoded_cache_training",
      "final_dir": "checkpoints/ablation-schema-v2-baseline8h4l-cache-10epoch-seed62/final",
      "data_file": "reports/dmhy_weak.template_generated.top5000.rust.schema_v2.candidate.jsonl",
      "dataset_mode": "encoded-cache-sharded",
      "tokenizer_variant": "char",
      "vocab_file": "datasets/AnimeName/vocab.char.json",
      "vocab_size": 6199,
      "label_schema_version": 2,
      "max_seq_length": 128,
      "hidden_size": 256,
      "num_hidden_layers": 4,
      "num_attention_heads": 8,
      "epochs": 10.0,
      "learning_rate": 4e-05,
      "warmup_steps": 120,
      "seed": 62,
      "role": "Base schema v2 checkpoint for repaired hard-focus training."
    },
    {
      "name": "ablation-schema-v2-hardfocus-cache-repaired-from-baseline-seed62-10epoch-rerun",
      "type": "schema_v2_repaired_hard_focus_encoded_cache_training",
      "final_dir": "checkpoints/ablation-schema-v2-hardfocus-cache-repaired-from-baseline-seed62-10epoch-rerun/final",
      "init_model_dir": "checkpoints/ablation-schema-v2-baseline8h4l-cache-10epoch-seed62/final",
      "data_file": "data/schema_v2_hard_focus_char_seed63.jsonl",
      "encoded_cache_dir": "data/encoded_cache/schema_v2_hard_focus_char_seed63_split995_repaired_v2",
      "dataset_mode": "encoded-cache-sharded",
      "tokenizer_variant": "char",
      "vocab_file": "datasets/AnimeName/vocab.char.json",
      "vocab_size": 6199,
      "label_schema_version": 2,
      "max_seq_length": 128,
      "hidden_size": 256,
      "num_hidden_layers": 4,
      "num_attention_heads": 8,
      "train_samples": 1068890,
      "eval_samples": 5372,
      "epochs": 10.0,
      "batch_size": 512,
      "learning_rate": 4e-05,
      "warmup_steps": 120,
      "seed": 63,
      "device": "cuda",
      "mixed_precision": "bf16",
      "tf32": true,
      "eval_f1": 0.9998627806136451,
      "eval_accuracy": 0.9999632255180959,
      "fixed_regression_model_only": "28/28",
      "fixed_regression_normalized_only": "28/28",
      "heldout_model_only": "2046/2048",
      "heldout_normalized_only": "2046/2048",
      "role": "Published repository-root checkpoint."
    }
  ],
  "release_reports": {
    "run_metadata": "reports/run_metadata.json",
    "trainer_eval_metrics": "reports/trainer_eval_metrics.json",
    "parse_eval_metrics": "reports/parse_eval_metrics.json",
    "case_metrics": "reports/case_metrics.json",
    "benchmark_results": "reports/benchmark_results.json",
    "onnx_metadata": "exports/anime_filename_parser.metadata.json"
  }
}