v3.2: 引号修复 + 数据扩标 + max_tokens→1024

README.md

@@ -7,13 +7,13 @@ tags: [qwen3, lora, narrative-parsing, novel-analysis, chinese-novels]
 base_model: mikuhhn1239/qwen3-8b-novel-base-sft
 ---
 
-# Qwen3-8B Narrative Type Classification LoRA
+# Qwen3-8B Narrative Type Classification LoRA (v3.2) ⭐
 
 **All Novel Can Be Galgame** — Agent 1: 叙事单元类型分类
 
 输入已切分的叙事单元，为每个 unit_id 标注类型。
 
-> 历史版本 v3 可通过 `revision="v3"` 加载。
+> 历史版本 v3 / v3.1 可通过 `revision="v3"` 和 `revision="v3.1"` 加载。
 
 基座: Qwen3-8B-Novel-Base-SFT | 方法: LoRA r=64 α=128 | 硬件: A800 80GB
 
@@ -22,13 +22,16 @@ base_model: mikuhhn1239/qwen3-8b-novel-base-sft
 ## 任务
 - 输入: 编号叙事单元 `[1] "..." [2] "..." ...`
 - 输出: `{"labels": [{"unit_id": "N", "type": "dialogue|narration|thought|action|scene_description"}]}`
+- 测试集: 39 条
 
 ### 五种类型
-- `dialogue` — 对话
-- `narration` — 叙述
-- `thought` — 心理
-- `action` — 动作
-- `scene_description` — 场景描写
+| 类型 | 含义 |
+|------|------|
+| `dialogue` | 对话 |
+| `narration` | 叙述 |
+| `thought` | 心理 |
+| `action` | 动作 |
+| `scene_description` | 场景描写 |
 
 ### 示例
 ```
@@ -55,7 +58,7 @@ base = AutoModelForCausalLM.from_pretrained(
     torch_dtype="auto", device_map="auto",
 )
 model = PeftModel.from_pretrained(
-    base, "mikuhhn1239/qwen3-8b-narrative-type-lora"
+    base, "mikuhhn1239/qwen3-8b-narrative-parsing-lora"
 )
 ```
 
@@ -63,24 +66,24 @@ model = PeftModel.from_pretrained(
 ```
 基座: Qwen3-8B-Novel-Base-SFT (Stage1 全参 SFT, 72K 小说续写数据)
 方法: LoRA (r=64, α=128, dropout=0.05)
-数据: 388 条 (310 train / 39 val / 39 test)，Clean 标注，预切分+人工标类型
+数据: 616 条 (577 train / 39 val / 39 test)
 框架: transformers Trainer + PEFT
 优化器: AdamW (adamw_torch_fused), cosine schedule, warmup=5%
 epoch: 5 | LR: 1e-4 | batch: 1×16(accum) | bf16 | max_length: 4096
 ```
 
 ### 版本历史
-
-| 版本 | 数据量 | epochs | LR | JSON解析 | 关键指标 |
-|------|:---:|:---:|:---:|:---:|:---:|
-| 零基座 (Qwen3-8B) | — | — | — | 0% | acc 0% |
-| +Stage1 (全参SFT) | — | — | — | 0% | acc 0% |
-| v1 | 70 | 3 | 2e-4 | 57.1% | acc 25.0% |
-| v2 | 310 | 3 | 2e-4 | 2.6% | acc 63.6% |
-| v3 (LoRA) | 310 | 3 | 2e-4 | 2.6% | acc 63.6% |
-| **v3.1 (LoRA)** | **310** | **5** | **1e-4** | **2.6%** | **acc 63.6%** |
+| 版本 | 数据量 | 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 | — | — | — | 0% | 0% | 读完 669 本也不会 |
+| v1 | 56 | 3 | 2e-4 | 57.1% | 25.0% | 端到端（切分+分类） |
+| v2 | 310 | 3 | 2e-4 | 2.6% | 63.6% | 只分类，引号冲突 |
+| v3.1 | 310 | 5 | 1e-4 | 2.6% | 63.6% | max_new_tokens=256 截断 JSON |
+| **v3.2** ⭐ | **577** | **5** | **1e-4** | **100%** | **69.5%** | **引号修复 + 扩标 + tokens→1024** |
 
 ### 结论
 
-- **零基座 / +Stage1 全 0%**：不做 Agent SFT 就不会叙事类型分类 ✅
-- **narrative-type**: acc 63.6% 但有意义，JSON 解析 2.6% 是格式问题（输入中原始引号干扰），修了就是有效模型
+- **零基座 / +Stage1 全 0%**：不做 Agent SFT 就不会叙事分类 ✅
+- **v3.2 突破**：JSON 解析 2.6%→100%，类型准确率 63.6%→69.5%（+5.9pp）
+- **关键修复**: 输入引号 `""`→`「」` + max_new_tokens 256→1024

adapter_config.json

@@ -1,48 +1,48 @@
-{
-  "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
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "mikuhhn1239/qwen3-8b-novel-base-sft",
+  "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": [
+    "up_proj",
+    "v_proj",
+    "o_proj",
+    "k_proj",
+    "q_proj",
+    "gate_proj",
+    "down_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:e1454b00f37e0dce8aacd61601de6982592bc83c9f4ccd9ce98a91b22fae1c40
+oid sha256:721979bc610ec0d5fb207cc553c4156e91b6714fd82b5e74237a37f5058fdca1
 size 698419728