Upload folder using huggingface_hub

README.md

@@ -3,43 +3,28 @@ language: zh
 license: apache-2.0
 library_name: peft
 pipeline_tag: text-generation
-tags:
-- qwen3
-- lora
-- narrative-parsing
-- novel-analysis
-- chinese-novels
-base_model: mikuhhn1239/qwen3-8b-novel-base-sft
+tags: [qwen3, lora, narrative-parsing, novel-analysis, chinese-novels]
+base_model: Qwen/Qwen3-8B-Instruct
 ---
 
-# Qwen3-8B Narrative Parsing LoRA
+# Qwen3-8B Narrative Type Classification LoRA
 
-**All Novel Can Be Galgame** 工作台 — Agent 1：叙事解析。
+**All Novel Can Be Galgame** — Agent 1: 叙事单元分类 (v3.1)
 
-将小说文本切分为叙事单元并标注类型。
+将已切分的叙事单元标注为 dialogue / narration / thought / action / scene_description。
 
-## 任务
+## 版本演进
 
-输入：中文小说片段  
-输出：`[{unit_id, text, type}]`
+| 版本 | 数据 | epochs | LR | JSON解析 | 类型准确率 | 说明 |
+|------|------|--------|-----|---------|-----------|------|
+| 零基座 | — | — | — | 0% | 0% | Qwen3-8B原始完全不会 |
+| +Stage1全参 | — | — | — | 0% | 0% | 读完669本小说也不会 |
+| v1 | 68 | 3 | 2e-4 | 14.3% | 6.2% | 端到端(切分+分类+复制)过难 |
+| v2 | ~140 | 3 | 2e-4 | 57.1% | 25.0% | 简化格式只分类 |
+| v3 | 310 | 3 | 2e-4 | 2.6% | **63.6%** | 清洁数据+统一Prompt |
+| **v3.1** | **310** | **5** | **1e-4** | **2.6%** | **63.6%** | **当前最佳** |
 
-五种类型：
-- `dialogue` — 角色对话
-- `narration` — 一般叙述
-- `thought` — 内心活动
-- `action` — 动作/行为
-- `scene_description` — 环境描写
-
-## 示例
-
-```
-输入: "你怎么来了？"她愣了一下。我站在门口，雨水顺着伞沿滴落。
-
-输出:
-  u1: "你怎么来了？"     → dialogue
-  u2: 她愣了一下。       → action
-  u3: 我站在门口，雨水... → narration
-```
+> 类型准确率 63.6% 但 JSON 解析 2.6% 是输出格式问题（引号/换行符转义），非判断错误。下版修复格式。
 
 ## 加载
 
@@ -47,30 +32,12 @@ base_model: mikuhhn1239/qwen3-8b-novel-base-sft
 from transformers import AutoModelForCausalLM
 from peft import PeftModel
 
-base = AutoModelForCausalLM.from_pretrained(
-    "mikuhhn1239/qwen3-8b-novel-base-sft",
-    torch_dtype="auto", device_map="auto",
-)
-model = PeftModel.from_pretrained(
-    base, "mikuhhn1239/qwen3-8b-narrative-parsing-lora"
-).merge_and_unload()
+base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B-Instruct", torch_dtype="auto", device_map="auto")
+model = PeftModel.from_pretrained(base, "mikuhhn1239/qwen3-8b-narrative-parsing-lora").merge_and_unload()
 ```
 
-## 版本历史
-
-| 版本 | JSON 解析率 | 类型准确率 | 比对单元 | 说明 |
-|------|-----------|-----------|---------|------|
-| v1 | 14.3% | 6.2% | 16 | 端到端（切分+分类+复制原文），68条数据 |
-| v2 | 57.1% | 25.0% | 92 | v2 简化格式（只分类），扩标数据 |
-| v3 | 待训 | 待训 | — | 388条清洁数据，统一System Prompt |
-
-> v1: 需同时做切分+分类+复制原文 → 过难
-> v2: 只输出 unit_id + type → JSON解析从14%→57%，类型准确从6%→25%
-> v3: 388条训练数据 + 更高质量标注 → 预期>70%类型准确率
-
 ## 训练
 
-- 基座：`mikuhhn1239/qwen3-8b-novel-base-sft`
-- 方法：LoRA (r=64, alpha=128, dropout=0.05)
-- 数据：15 本标注小说的 68 条训练样本 (v1 格式)
-- 硬件：A800
+- 基座: Qwen3-8B-Instruct | 方法: LoRA r=64 α=128
+- 数据: 388条 (310 train / 39 val / 39 test)
+- 格式: 只输出 `{"labels": [{"unit_id","type"}]}`，不复制原文

adapter_config.json

@@ -1,48 +1,48 @@
-{
-  "alora_invocation_tokens": null,
-  "alpha_pattern": {},
-  "arrow_config": null,
-  "auto_mapping": null,
-  "base_model_name_or_path": "/workspace/project-nas-1000073/linyupeng/data/checkpoints/stage1-base-sft/final",
-  "bias": "none",
-  "corda_config": null,
-  "ensure_weight_tying": false,
-  "eva_config": null,
-  "exclude_modules": null,
-  "fan_in_fan_out": false,
-  "inference_mode": true,
-  "init_lora_weights": true,
-  "layer_replication": null,
-  "layers_pattern": null,
-  "layers_to_transform": null,
-  "loftq_config": {},
-  "lora_alpha": 128,
-  "lora_bias": false,
-  "lora_dropout": 0.05,
-  "lora_ga_config": null,
-  "megatron_config": null,
-  "megatron_core": "megatron.core",
-  "modules_to_save": null,
-  "peft_type": "LORA",
-  "peft_version": "0.19.1",
-  "qalora_group_size": 16,
-  "r": 64,
-  "rank_pattern": {},
-  "revision": null,
-  "target_modules": [
-    "k_proj",
-    "o_proj",
-    "up_proj",
-    "v_proj",
-    "q_proj",
-    "down_proj",
-    "gate_proj"
-  ],
-  "target_parameters": null,
-  "task_type": "CAUSAL_LM",
-  "trainable_token_indices": null,
-  "use_bdlora": null,
-  "use_dora": false,
-  "use_qalora": false,
-  "use_rslora": false
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-8B-Instruct",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 128,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.1",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k_proj",
+    "gate_proj",
+    "up_proj",
+    "down_proj",
+    "o_proj",
+    "q_proj",
+    "v_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
 }

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0991c5b27e8772ad9781a30560ab7eb5a3400f2f1839e413cd86db0f20a4fd2f
+oid sha256:e1454b00f37e0dce8aacd61601de6982592bc83c9f4ccd9ce98a91b22fae1c40
 size 698419728

tokenizer.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c0acdaba32b920d640afb36af4396c91974e074735636e4016d17a8ed9c03730
+oid sha256:2f1298e298f2fe0059aba46f037697a339ccba45a1908780ce8ca14b45582f23
 size 11422753