DocTalk
收藏资源简介:
DocTalk是一个多轮预训练对话语料库,包含超过730k长的对话,由南洋理工大学和亚马逊的研究人员创建。该数据集通过将多个相关文档转换为一个扩展的多轮、多主题信息寻求对话来合成会话数据。数据集的大小为730,707个对话,每个对话平均包含82.2个轮次,其中助手的话语平均长度为87.3个单词,用户的话语平均长度为26.5个单词。数据集创建过程分为三个阶段:文档图构建、对话图构建和用户话语生成。DocTalk旨在解决大型语言模型在多轮对话任务中的能力与训练范式之间的潜在不匹配问题,通过在预训练期间接触合成对话结构来增强LLM的基本多轮能力,如上下文记忆和理解。DocTalk适用于会话AI助手,以提供动态和自然的对话,并解决现实世界交互中的多轮、多主题对话需求。
DocTalk is a multi-turn pre-trained dialogue corpus created by researchers from Nanyang Technological University and Amazon, comprising 730,707 long conversations in total. On average, each conversation contains 82.2 turns, with the assistant's utterances averaging 87.3 words per turn and the user's utterances averaging 26.5 words per turn. This dataset synthesizes conversational data by converting multiple related documents into extended multi-turn, multi-topic information-seeking dialogues. The construction process of DocTalk is divided into three stages: document graph construction, dialogue graph construction, and user utterance generation. DocTalk aims to address the potential mismatch between the capabilities and training paradigms of large language models (LLMs) in multi-turn dialogue tasks, and to enhance the core multi-turn capabilities of LLMs such as contextual memory and understanding by exposing the models to synthetic dialogue structures during pre-training. DocTalk is applicable to conversational AI assistants, enabling them to provide dynamic and natural dialogues and addressing the demand for multi-turn, multi-topic dialogues in real-world human-AI interactions.
数据集概述:DocTalk
基本信息
- 名称: DocTalk
- 许可证: CC-BY-SA-3.0
- 任务类别: 文本生成、问答、文本到文本生成
- 语言: 英语
- 规模: 100K<n<1M
- 论文链接: https://arxiv.org/abs/2507.05750
数据集描述
DocTalk是一个大规模合成对话语料库,通过三阶段流程将相关的Wikipedia文档集群转化为多轮、多主题的信息寻求对话。
生成流程
- 文档图构建: 通过加权随机游走在有向无环图上采样最多三篇相关的Wikipedia文章。
- 对话图构建: 将每篇文章分段,构建完全连接的段落图,并通过微调的对话奖励模型采样助手话语。
- 用户话语生成: 使用Mistral-2-7B-Instruct模型生成用户问题。
关键统计
- 对话总数: 730,707
- 每对话平均轮数: 82.2
- 助手话语平均长度: 87.3词
- 用户话语平均长度: 26.5词
- 每对话平均文档转换次数: 23.3
数据集结构
每个样本为JSON对象,包含:
- id: 唯一对话标识符
- dialog: 对话轮次列表,每轮包含:
- role: "user"或"assistant"
- src_document: Wikipedia文章标题
- text: 话语文本
- turn: 对话轮次索引
样本示例
基于三篇Wikipedia文章的对话示例:
- 用户: 多伦多总医院位于何处?有何特点?
- 助手: 多伦多总医院是加拿大安大略省多伦多市的一家主要教学医院...
- 用户: 多伦多大学医学院的医学博士项目后两年涵盖哪些专业?
下载方式
python from datasets import load_dataset dataset = load_dataset("AmazonScience/DocTalk")
引用
bibtex @misc{lee2025doctalkscalablegraphbaseddialogue, title={DocTalk: Scalable Graph-based Dialogue Synthesis for Enhancing LLM Conversational Capabilities}, author={Jing Yang Lee and Hamed Bonab and Nasser Zalmout and Ming Zeng and Sanket Lokegaonkar and Colin Lockard and Binxuan Huang and Ritesh Sarkhel and Haodong Wang}, year={2025}, eprint={2507.05750}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2507.05750}, }




