| { |
| "schema_version": 2, |
| "notebook": "SFT_20260711_run_1.ipynb", |
| "run_name": "qwen3-4b-structured-output-lora_0711_run_1", |
| "experimental_question": "2026-07-10 Run 3のYAML中心ベース＋CSV→JSONリプレイを維持し、YAML→TOML教師トークンを最終混合の約10%追加することで、JSON／CSVの改善を保ちながらTOML出力を改善できるか。", |
| "public_benchmark_policy": "public_150.jsonは読み込まず、学習、追加データ選択、チェックポイント選択、早期終了のいずれにも使用しません。", |
| "base_model": "Qwen/Qwen3-4B-Instruct-2507", |
| "max_seq_length": 2048, |
| "datasets": { |
| "base": { |
| "id": "daichira/structeval-t-sft-hq-yaml-cleaned", |
| "revision": "e73a755d3b50bd3ac38c99c1ede87ca9b3139cba" |
| }, |
| "replay": { |
| "id": "daichira/structured-5k-mix-sft", |
| "revision": "740e82f5bb5ae31c9b0f9c2fa15f78bbcaf3dbe2", |
| "filter": { |
| "category": "C_JSON", |
| "subcategory": "csv_to_json", |
| "task": "transform" |
| } |
| }, |
| "toml_replay": { |
| "id": "daichira/structured-5k-mix-sft", |
| "revision": "740e82f5bb5ae31c9b0f9c2fa15f78bbcaf3dbe2", |
| "target": { |
| "category": "C_TOML", |
| "subcategory": "yaml_to_toml", |
| "task": "transform" |
| }, |
| "primary_source": { |
| "category": "C_YAML", |
| "subcategory": "toml_to_yaml", |
| "task": "transform", |
| "derivation": "Reverse validated TOML→YAML pairs." |
| }, |
| "fallback_source": { |
| "category": "C_TOML", |
| "subcategory": "text_to_toml", |
| "task": "extract", |
| "derivation": "Parse target TOML and deterministically serialize the same object as YAML input." |
| } |
| } |
| }, |
| "seeds": { |
| "split_seed": 3407, |
| "lora_seed": 3407, |
| "trainer_seed": 42, |
| "data_seed": 42, |
| "toml_selection_seed": 1042 |
| }, |
| "optimizer_trajectory": { |
| "training_budget_mode": "epochs", |
| "requested_num_train_epochs": 5.0, |
| "requested_max_steps": null, |
| "warmup_mode": "ratio", |
| "warmup_ratio": 0.1, |
| "learning_rate": 1e-06, |
| "scheduler": "cosine", |
| "weight_decay": 0.05, |
| "micro_batch": 2, |
| "gradient_accumulation": 8, |
| "effective_batch": 16 |
| }, |
| "replay_selection": { |
| "target_replay_supervised_tokens": 55774, |
| "selected_replay_supervised_tokens": 56009, |
| "target_mixed_token_share": 0.1, |
| "actual_mixed_token_share": 0.1003795889399468, |
| "selected_rows": 79, |
| "available_rows": 85, |
| "pool_exhausted_before_target": false, |
| "selected_length_bins": { |
| "long": 28, |
| "medium": 29, |
| "short": 22 |
| }, |
| "selected_root_types": { |
| "object": 79 |
| } |
| }, |
| "toml_replay_selection": { |
| "target_toml_supervised_tokens": 61997, |
| "selected_toml_supervised_tokens": 62341, |
| "target_final_token_share": 0.1, |
| "actual_final_token_share": 0.10049926408119772, |
| "selected_rows": 104, |
| "primary_selection": { |
| "target_supervised_tokens": 61997, |
| "selected_supervised_tokens": 62341, |
| "selected_rows": 104, |
| "available_rows": 193, |
| "pool_exhausted_before_target": false, |
| "selected_length_bins": { |
| "short": 60, |
| "medium": 30, |
| "long": 14 |
| }, |
| "selected_origins": { |
| "reversed_toml_to_yaml": 104 |
| } |
| }, |
| "fallback_selection": { |
| "target_supervised_tokens": 0, |
| "selected_supervised_tokens": 0, |
| "selected_rows": 0, |
| "available_rows": 1483, |
| "pool_exhausted_before_target": false, |
| "selected_length_bins": {} |
| }, |
| "pool_exhausted_before_target": false, |
| "length_cycle": [ |
| "short", |
| "short", |
| "short", |
| "short", |
| "medium", |
| "medium", |
| "long" |
| ], |
| "selected_origins": { |
| "reversed_toml_to_yaml": 104 |
| } |
| }, |
| "final_supervised_token_composition": { |
| "base_yaml_centric": { |
| "tokens": 501963, |
| "final_share": 0.8092092217960932 |
| }, |
| "csv_to_json": { |
| "tokens": 56009, |
| "final_share": 0.09029151412270901 |
| }, |
| "yaml_to_toml": { |
| "tokens": 62341, |
| "final_share": 0.10049926408119772 |
| } |
| }, |
| "token_statistics": { |
| "base_train": { |
| "rows": 1900, |
| "full_fit_rows": 1806, |
| "full_fit_rate": 0.9505263157894737, |
| "all_masked_rows": 0, |
| "visible_supervised_tokens": 501963, |
| "mean_prompt_retention": 0.9936397172610149, |
| "mean_full_tokens": 568.43, |
| "p50_full_tokens": 228, |
| "p90_full_tokens": 1803, |
| "max_full_tokens": 2495 |
| }, |
| "base_validation": { |
| "rows": 100, |
| "full_fit_rows": 92, |
| "full_fit_rate": 0.92, |
| "all_masked_rows": 0, |
| "visible_supervised_tokens": 24865, |
| "mean_prompt_retention": 0.9890263754525934, |
| "mean_full_tokens": 535.09, |
| "p50_full_tokens": 239, |
| "p90_full_tokens": 1700, |
| "max_full_tokens": 2403 |
| }, |
| "json_replay_train_pool": { |
| "rows": 85, |
| "full_fit_rows": 85, |
| "full_fit_rate": 1.0, |
| "all_masked_rows": 0, |
| "visible_supervised_tokens": 58991, |
| "mean_prompt_retention": 1.0, |
| "mean_full_tokens": 1364.0, |
| "p50_full_tokens": 1392, |
| "p90_full_tokens": 1852, |
| "max_full_tokens": 1945 |
| }, |
| "selected_json_replay_train": { |
| "rows": 79, |
| "full_fit_rows": 79, |
| "full_fit_rate": 1.0, |
| "all_masked_rows": 0, |
| "visible_supervised_tokens": 56009, |
| "mean_prompt_retention": 1.0, |
| "mean_full_tokens": 1393.113924050633, |
| "p50_full_tokens": 1401, |
| "p90_full_tokens": 1861, |
| "max_full_tokens": 1945 |
| }, |
| "json_replay_dev": { |
| "rows": 10, |
| "full_fit_rows": 10, |
| "full_fit_rate": 1.0, |
| "all_masked_rows": 0, |
| "visible_supervised_tokens": 6859, |
| "mean_prompt_retention": 1.0, |
| "mean_full_tokens": 1354.1, |
| "p50_full_tokens": 1352, |
| "p90_full_tokens": 1820, |
| "max_full_tokens": 1999 |
| }, |
| "toml_replay_train_pool": { |
| "rows": 1676, |
| "full_fit_rows": 1676, |
| "full_fit_rate": 1.0, |
| "all_masked_rows": 0, |
| "visible_supervised_tokens": 232041, |
| "mean_prompt_retention": 1.0, |
| "mean_full_tokens": 341.563245823389, |
| "p50_full_tokens": 201, |
| "p90_full_tokens": 1001, |
| "max_full_tokens": 2019 |
| }, |
| "selected_toml_replay_train": { |
| "rows": 104, |
| "full_fit_rows": 104, |
| "full_fit_rate": 1.0, |
| "all_masked_rows": 0, |
| "visible_supervised_tokens": 62341, |
| "mean_prompt_retention": 1.0, |
| "mean_full_tokens": 1280.2403846153845, |
| "p50_full_tokens": 1239, |
| "p90_full_tokens": 1696, |
| "max_full_tokens": 2002 |
| }, |
| "toml_replay_dev": { |
| "rows": 20, |
| "full_fit_rows": 20, |
| "full_fit_rate": 1.0, |
| "all_masked_rows": 0, |
| "visible_supervised_tokens": 3482, |
| "mean_prompt_retention": 1.0, |
| "mean_full_tokens": 414.05, |
| "p50_full_tokens": 202, |
| "p90_full_tokens": 1428, |
| "max_full_tokens": 2044 |
| }, |
| "mixed_train": { |
| "rows": 2083, |
| "full_fit_rows": 1989, |
| "full_fit_rate": 0.9548727796447432, |
| "all_masked_rows": 0, |
| "visible_supervised_tokens": 620313, |
| "mean_prompt_retention": 0.99419849390107, |
| "mean_full_tokens": 635.2462794047048, |
| "p50_full_tokens": 250, |
| "p90_full_tokens": 1797, |
| "max_full_tokens": 2495 |
| } |
| }, |
| "row_counts": { |
| "base_raw": 2000, |
| "base_train": 1900, |
| "base_validation": 100, |
| "structured_mix_raw": 5000, |
| "json_replay_clean_before_token_filter": 100, |
| "json_replay_full_fit": 95, |
| "json_replay_train_pool": 85, |
| "selected_json_replay_train": 79, |
| "json_replay_dev": 10, |
| "toml_primary_clean_before_token_filter": 200, |
| "toml_fallback_clean_before_token_filter": 1500, |
| "toml_primary_full_fit": 196, |
| "toml_fallback_full_fit": 1500, |
| "toml_replay_train_pool": 1676, |
| "selected_toml_replay_train": 104, |
| "toml_replay_dev": 20, |
| "mixed_train": 2083 |
| }, |
| "replay_rejection_reasons": { |
| "csv_to_json": { |
| "wrong_category": 4500, |
| "wrong_subcategory": 400 |
| }, |
| "yaml_to_toml": {} |
| }, |
| "replay_token_filter_rejections": { |
| "csv_to_json": 5, |
| "yaml_to_toml_primary": 4, |
| "yaml_to_toml_fallback": 0 |
| }, |
| "id_hashes": { |
| "base_train_ids_sha256": "5a572c3d63033e85893a2516892f503c696f1e1dae61fae614722ac615dc6a01", |
| "base_validation_ids_sha256": "286e8d907a1e2ae51708f10f8d4ce164b4e7f1ebba21f8e756cc864dc6cad9ac", |
| "json_replay_train_ids_sha256": "9e7d02e2fcf7d560fe0b217686f2ceedb5a5fc9229494d373424f324a307bd0f", |
| "json_replay_dev_ids_sha256": "7bc831d37d3053ddcaa9691779b01163a677461767a3b4de44e6fc4f92a5b25a", |
| "toml_replay_train_ids_sha256": "dea099f91255232d1f1a76c96cb7add983651161023467818c4ce78e1d053742", |
| "toml_replay_dev_ids_sha256": "1b751d55f2618f30d7182390ad99d61af4eae9729241389bf847334cc6af43d8" |
| }, |
| "tokenizer": { |
| "model_id": "Qwen/Qwen3-4B-Instruct-2507", |
| "chat_template_sha256": "64f85b198065d0fba2a81f37e10ed68161ce2c19a754c7100e67e0ca2ee9c326", |
| "vocab_size": 151669 |
| }, |
| "prepared_paths": { |
| "base_train": "/content/SFT_20260711_run_1_prepared/base_train", |
| "base_val": "/content/SFT_20260711_run_1_prepared/base_val", |
| "json_replay_train": "/content/SFT_20260711_run_1_prepared/json_replay_train", |
| "json_replay_dev": "/content/SFT_20260711_run_1_prepared/json_replay_dev", |
| "toml_replay_train": "/content/SFT_20260711_run_1_prepared/toml_replay_train", |
| "toml_replay_dev": "/content/SFT_20260711_run_1_prepared/toml_replay_dev", |
| "mixed_train": "/content/SFT_20260711_run_1_prepared/mixed_train" |
| } |
| } |
