PubMed多模型医学推理数据集

本数据集来自 PubMed，共包含 940 个 case。对于同一病例，使用多款推理/诊断模型（deepseek-r1、o3-mini、gemini2-ft、qwq、baichuan-m1、diagnosegpt、medgemma）分别进行多轮或自由轮次的问答推理，输出分为 reasoning（推理过程 / 中间思路）和 answer（最终结论 / 建议）。数据用于模型性能对比、推理可解释性分析与临床问答研究（仅作科研与开发用途，不作为临床诊断依据）。给数据原始出处来自HuggingFace[Henrychur/MedRbench-Inference-Results]，这个数据集主要将其中内容翻译成中文内容，方便使用。

This dataset is sourced from PubMed and contains a total of 940 clinical cases. For each individual case, multiple inference/diagnostic models (deepseek-r1, o3-mini, gemini2-ft, qwq, baichuan-m1, diagnosegpt, medgemma) were used to conduct multi-round or open-ended question-answering and reasoning separately. The model outputs are divided into two parts: reasoning (reasoning process / intermediate thinking steps) and answer (final conclusion / clinical recommendations). This dataset is intended for model performance comparison, reasoning interpretability analysis and clinical question-answering research, and is for scientific research and development purposes only, not to be used as a basis for clinical diagnosis. The original source of the core content of this dataset is Hugging Face [Henrychur/MedRbench-Inference-Results]; this specific version primarily translates the original content into Chinese for easier usage.

搜集汇总

背景与挑战

背景概述

该数据集源自PubMed，包含940个医学案例，每个案例记录了多个推理模型（如deepseek-r1、o3-mini等）的对话输出，其中包含推理过程和最终答案。数据用于模型性能比较和医学推理研究，但仅限研发用途，不应用于临床诊断。