MedReason
收藏资源简介：
MedReason是一个大规模高质量的医学推理数据集，旨在使大型语言模型（LLMs）能够进行忠实和可解释的医学问题解决。我们利用结构化的医学知识图（KG）将临床问答对转换为逻辑推理链或“思考路径”。我们的管道为来自7个医学数据集的各种医学问题生成详细的推理，结果是一个包含32,682个问答对的数据集，每个问答对都有详细的逐步解释。
MedReason is a large-scale, high-quality medical reasoning dataset aimed at enabling large language models (LLMs) to conduct faithful and interpretable medical problem-solving. We leverage structured medical knowledge graphs (KGs) to transform clinical question-answer pairs into logical reasoning chains or "thinking paths". Our pipeline generates detailed inferences for various medical questions sourced from 7 medical datasets, resulting in a dataset containing 32,682 question-answer pairs, each accompanied by a comprehensive step-by-step explanation.
MedReason数据集概述
数据集简介
- 名称: MedReason
- 类型: 医疗推理数据集
- 规模: 32,682个问答对
- 特点:
- 基于结构化医学知识图谱(KG)构建
- 提供详细的逐步解释推理链
- 覆盖7个医疗数据集的问题
数据内容
- 数据生成方法:
- 使用临床问答对和知识图谱生成思维链推理
- 生成代码位于
./src/data_generation
- 数据访问:
- 质量过滤后的监督微调(SFT)数据
- 访问地址: https://huggingface.co/datasets/UCSC-VLAA/MedReason
相关模型
- MedReason-8B:
- 基础模型: HuatuoGPT-o1-8B
- 访问地址: https://huggingface.co/UCSC-VLAA/MedReason-8B
- MedReason-Llama:
- 基础模型: Llama-3.1-8B-Instruct
- 访问地址: https://huggingface.co/UCSC-VLAA/MedReason-Llama
- MedReason-Mistral:
- 基础模型: Mistral-7B-Instruct-v0.2
- 访问地址: https://huggingface.co/UCSC-VLAA/MedReason-Mistral
训练与评估
- 训练方法:
- 使用8-GPU进行监督微调(SFT)
- 提供基于Huatuo-o1-8B和DeepSeek-distilled-Llama-8B的训练脚本
- 评估结果:
- 在医疗基准测试中表现优异
- 提供案例研究和定量结果展示
引用信息
bibtex @misc{wu2025medreasonelicitingfactualmedical, title={MedReason: Eliciting Factual Medical Reasoning Steps in LLMs via Knowledge Graphs}, author={Juncheng Wu and Wenlong Deng and Xingxuan Li and Sheng Liu and Taomian Mi and Yifan Peng and Ziyang Xu and Yi Liu and Hyunjin Cho and Chang-In Choi and Yihan Cao and Hui Ren and Xiang Li and Xiaoxiao Li and Yuyin Zhou}, year={2025}, eprint={2504.00993}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2504.00993}, }
