Document training lineage

README.md

@@ -54,6 +54,7 @@ This repository is the Hugging Face model repo used by MiruPlay as `tools/anime_
 | Labels / 标签 | BIO labels for `TITLE`, `SEASON`, `EPISODE`, `GROUP`, `RESOLUTION`, `SOURCE`, `SPECIAL` |
 | Default checkpoint / 默认权重 | Repository root files (`config.json`, `model.safetensors`, `vocab.json`, `tokenizer_config.json`) |
 | ONNX export / ONNX 导出 | `exports/anime_filename_parser.onnx` |
+| Training lineage / 训练链路 | `training_lineage.json` |
 
 **中文**：根目录就是发布 checkpoint，不再保留旧的 `model/` 重复副本。默认解析路径是“模型 logits + 约束 BIO + 薄字段规范化”，不再默认启用重结构规则；直接 `from_pretrained()` 只能加载 token-classification 权重。
 
@@ -147,9 +148,9 @@ Current published checkpoint:
 | ONNX parity / ONNX 误差 | max abs diff `4.0531e-05` |
 | CPU thin-runtime latency / CPU 薄层运行时延迟 | ONNX avg `13.08 ms`, P95 `15.95 ms` |
 
-**中文**：当前发布模型是“全量重标注 char 模型 + thin hard-case focus 微调”。README 主指标以 `model-only` 和默认薄层 `normalized-only` 为准；`--rule-assist` 只保留为兼容/诊断对照，不再作为模型质量标准。
+**中文**：当前发布模型是“两阶段训练”产物：先在 `datasets/AnimeName/dmhy_weak_char.jsonl` 上全量 CUDA 重训，再做 thin hard-case focus 微调。细节见 `training_lineage.json`。README 主指标以 `model-only` 和默认薄层 `normalized-only` 为准；`--rule-assist` 只保留为兼容/诊断对照，不再作为模型质量标准。
 
-**English**: The published checkpoint is the full-relabel character model plus a thin hard-case focus fine-tune. README quality numbers prioritize `model-only` and the default thin `normalized-only` runtime; `--rule-assist` is retained only for compatibility/diagnostics.
+**English**: The published checkpoint was trained in two stages: a full CUDA fine-tune on `datasets/AnimeName/dmhy_weak_char.jsonl`, followed by a thin hard-case focus fine-tune. See `training_lineage.json` for details. README quality numbers prioritize `model-only` and the default thin `normalized-only` runtime; `--rule-assist` is retained only for compatibility/diagnostics.
 
 Run regression:
 

training_lineage.json

@@ -0,0 +1,55 @@
+{
+  "published_checkpoint": "repository_root",
+  "summary": "The published checkpoint was produced in two stages: a full-dataset CUDA fine-tune on dmhy_weak_char.jsonl, followed by a thin-runtime hard-case focus fine-tune.",
+  "summary_zh": "当前发布 checkpoint 是两阶段产物：先在 dmhy_weak_char.jsonl 上做全量 CUDA 微调，再做薄层运行时困难样本微调。",
+  "stages": [
+    {
+      "name": "dmhy-char-thin-gpu",
+      "type": "full_dataset_finetune",
+      "machine": "adqew@192.168.63.157",
+      "data_file": "datasets/AnimeName/dmhy_weak_char.jsonl",
+      "tokenizer_variant": "char",
+      "vocab_file": "datasets/AnimeName/vocab.char.json",
+      "vocab_size": 6199,
+      "max_seq_length": 128,
+      "train_samples": 619361,
+      "eval_samples": 12641,
+      "epochs": 2.0,
+      "batch_size": 256,
+      "learning_rate": 0.00006,
+      "warmup_steps": 300,
+      "seed": 55,
+      "device": "cuda",
+      "fp16": true,
+      "eval_f1": 0.9962419041019217,
+      "eval_accuracy": 0.9991988685517916,
+      "role": "Base checkpoint for the final hard-case focus stage."
+    },
+    {
+      "name": "dmhy-char-thin-hardfocus",
+      "type": "hard_case_focus_finetune",
+      "machine": "adqew@192.168.63.157",
+      "data_file": "data/thin_hard_focus_char.jsonl",
+      "tokenizer_variant": "char",
+      "vocab_file": "datasets/AnimeName/vocab.char.json",
+      "vocab_size": 6199,
+      "max_seq_length": 128,
+      "train_samples": 117089,
+      "eval_samples": 6163,
+      "epochs": 2.0,
+      "batch_size": 256,
+      "learning_rate": 0.00004,
+      "warmup_steps": 80,
+      "seed": 58,
+      "device": "cuda",
+      "fp16": true,
+      "eval_f1": 0.9972066016906769,
+      "eval_accuracy": 0.9994733938512463,
+      "fixed_regression_model_only": "25/26",
+      "fixed_regression_normalized_only": "26/26",
+      "heldout_model_only": "1014/1024",
+      "heldout_normalized_only": "1017/1024",
+      "role": "Published repository-root checkpoint."
+    }
+  ]
+}