mkurman/medmcqa-hard

MedMCQA-Hard是一个更难、去重版本的MedMCQA，旨在减少记忆依赖并加强医学多项选择题的泛化能力。每个正确选项以多种表述/列表变体出现，使得模型不能依赖于表面形式的回忆，而必须对内容进行推理。每个项目包含一个标准正确答案和单一或一组错误答案，便于DPO、RLAIF/GRPO和对比目标的使用。此外，数据集还加入了轻量级的消息格式，方便指令调整模型和SFT的使用。

MedMCQA-Hard is a harder, de-duplicated version of MedMCQA designed to reduce memorization and enhance generalization of medical multiple-choice questions. Each correct option appears in multiple phrasing/list variants, preventing models from relying on surface-form recall and necessitating reasoning over content. Every item includes one canonical correct answer and both a single and a set of incorrect answers, making it suitable for DPO, RLAIF/GRPO, and contrastive objectives. The dataset also features lightweight message formatting, convenient for instruction-tuned models and SFT.