novel-agent-sft-dataset
收藏资源简介:
该数据集名为“All Novel Can Be Galgame — Agent SFT 训练数据”,是一个专门用于中文小说叙事理解 Agent SFT(监督微调)训练的数据集,旨在支持 Qwen3-8B 模型的 LoRA(低秩适应)微调。数据集包含多个迭代版本(从 base-sft 到 v4.1),数据量从 72K 条到 2,123 条不等,覆盖三个核心 Agent 任务:叙事类型分类(输入为已切分的叙事单元,输出为带单元 ID 和类型的标签)、场景边界检测(输入为编号段落序列,输出为边界位置列表)以及角色归因(输入为候选角色及上下文,输出为最佳候选角色和不确定性标志)。数据以 ChatML messages 格式([system, user, assistant])组织,存储为 JSONL 文件,可直接与 LLaMA-Factory 训练框架兼容。数据集的开发注重版本优化,如 v4 版本通过 DeepSeek 重标注在场景边界检测任务上首次实现 F1 分数突破 30%,适用于自然语言处理中的小说文本分析、叙事结构理解和角色属性推断等任务。
The dataset is named All Novel Can Be Galgame — Agent SFT Training Data and is specifically designed for training Agent SFT (Supervised Fine-Tuning) for Chinese novel narrative understanding, aiming to support LoRA (Low-Rank Adaptation) fine-tuning of the Qwen3-8B model. It includes multiple iterative versions (from base-sft to v4.1), with data volumes ranging from 72K entries to 2,123 entries, covering three core Agent tasks: narrative type classification (input is segmented narrative units, output is labels with unit IDs and types), scene boundary detection (input is a sequence of numbered paragraphs, output is a list of boundary positions), and character attribution (input is candidate characters and context, output is the best candidate character and an uncertainty flag). The data is organized in ChatML messages format ([system, user, assistant]) and stored as JSONL files, making it directly compatible with the LLaMA-Factory training framework. The dataset development focuses on version optimization, such as the v4 version achieving an F1 score breakthrough of over 30% in scene boundary detection through DeepSeek re-labeling. It is suitable for tasks in natural language processing, including novel text analysis, narrative structure understanding, and character attribute inference.
数据集概览
- 数据集名称: All Novel Can Be Galgame — Complete Dataset
- 语言: 中文 (zh)
- 许可证: Apache 2.0
- 任务类别: 文本生成、文本分类
- 标签: 中文小说、叙事分析、场景分割、角色归因、Agent SFT、Qwen3、LoRA
- 数据规模: 1M < n < 10M
- 总大小: 2.2 GB
- 总文件数: 42,191
数据规模详情
| 目录 | 文件数 | 大小 | 说明 |
|---|---|---|---|
training/ |
52 | 689 MB | 训练用 SFT 数据 (JSONL) |
raw-books/ |
671 | 327 MB | 669 本原始小说 |
processed/ |
39,842 | 1.2 GB | 按章节预处理文本 |
annotations/ |
1,626 | 1 MB | 原始标注文件 |
| 合计 | 42,191 | 2.2 GB |
目录结构
datasets/ ├── training/ │ ├── base-sft/ Stage1 全参 SFT (72K) │ ├── v1-attribution-assist/ v1 角色归因 (66 条) │ ├── v1-narrative-parsing/ v1 叙事解析 (70 条) │ ├── v1-scene-segmentation/ v1 场景分割 (82 条) │ ├── v2-attribution-best-candidate/ v2 角色归因 (300 条) │ ├── v2-narrative-type-classification/ v2 叙事分类 (388 条) │ ├── v2-scene-boundary-detection/ v2 场景边界 (~200 条) │ ├── v3.1-attribution-best-candidate/ v3.1 角色归因 (300 条) │ ├── v3.1-narrative-type-classification/v3.1 叙事分类 (388 条) │ ├── v3.1-scene-boundary-detection/ v3.1 场景边界 (315 条) │ ├── v3.2-attribution-best-candidate/ v3.2 角色归因 (525 条) ⭐ │ ├── v3.2-narrative-type-classification/v3.2 叙事分类 (655 条) ⭐ │ ├── v3.2-scene-boundary-detection/ v3.2 场景边界 (454 条) │ ├── v3.3-scene-boundary-detection/ v3.3 二元决策实验 (349 条) │ ├── v3.4-scene-boundary-detection/ v3.4 滑动窗口实验 (4505 条) │ ├── v4-scene-boundary-detection/ v4 DeepSeek重标注 (2123 条) ⭐⭐ │ └── v4.1-scene-boundary-detection/ v4.1 扩标 (685 条) ├── raw-books/ # 669 本中文小说原文 ├── processed/ # 预处理后的章节文本 └── annotations/ # 原始标注 JSON ├── attribution-assist/ 角色归因 ├── narrative-parsing/ 叙事解析 ├── scene-segmentation/ 场景分割 └── v2-expand/ 扩展标注
三个 Agent 任务
| Agent | 任务 | 输入 | 输出 |
|---|---|---|---|
| Agent 1 | 叙事类型分类 | 已切分叙事单元 | {labels: [{unit_id, type}]} |
| Agent 2 | 场景边界检测 | 编号段落序列 | {boundaries: [N], reasons: [...]} |
| Agent 3 | 角色归因 | 候选角色 + 上下文 | {best_candidate, uncertain} |
训练结果
| 版本 | 数据量 | narrative-type (acc) | scene-boundary (F1) | attr-best (acc) |
|---|---|---|---|---|
| 零基座 (Qwen3-8B) | — | 0% | 0% | 0% |
| v1 | 218 条 | 25.0% | 33.3% | 14.3% |
| v2 | 700+ 条 | 63.6% | 53.3% ⭐ | 33.3% |
| v3.1 | 900+ 条 | 63.6% | 28.6% | 43.3% |
| v3.2 | 1,634 条 | 69.5% ⭐ | 20.0% | 86.7% ⭐ |
| v3.3–v3.4 | 4,854 条 | — | 12–15% | — |
| v4 | 2,123 条 | — | 30.5% ⭐⭐ | — |
数据格式
训练数据为 JSONL 格式,每行一条 ChatML messages:
json { "messages": [ {"role": "system", "content": "系统提示词"}, {"role": "user", "content": "输入文本"}, {"role": "assistant", "content": "{"labels": [...]}"} ] }
相关模型
- 基座: Qwen/Qwen3-8B-Instruct
- Stage1 SFT: mikuhhn1239/qwen3-8b-novel-base-sft
- Agent 1: mikuhhn1239/qwen3-8b-narrative-parsing-lora
- Agent 2: mikuhhn1239/qwen3-8b-scene-segmentation-lora
- Agent 3: mikuhhn1239/qwen3-8b-attribution-assist-lora




