disi-unibo-nlp/MedQAbstain
收藏资源简介:
MedQAbstain是一个用于评估大语言模型(LLMs)在医学不确定性下弃权(abstention)能力的基准数据集。它通过改造现有的医学多项选择题(MCQA)数据集,移除原始正确答案并添加明确的我弃权选项(例如升级给其他医生),从而将问题转化为仅允许弃权的决策问题。该数据集旨在测试模型在缺乏安全行动选择时是否能够避免采取不安全行动。数据集包含两个子集:Safe(安全)和LT(Life-Threatening,生命威胁),分别对应不同临床风险级别。支持文本和多模态输入,并提供两种弃权设置:Trivial Abstention(隐藏问题,仅显示选项)和Standard Abstention(显示完整问题和选项)。此外,还提供针对推理型和非推理型模型的提示变体。数据集字段包括ID、来源数据集、问题文本、选项、答案索引、图像、模态类型、原始选项、原始答案以及多种提示字段。
MedQAbstain is a benchmark dataset for evaluating the abstention capability of Large Language Models (LLMs) in the context of medical uncertainty. It is constructed by modifying existing medical multiple-choice question answering (MCQA) datasets: removing the original correct answer and adding an explicit "I abstain" option (e.g., "consult another physician"), thereby transforming the original task into a decision-making problem where abstention is the only permissible action. The core goal of this dataset is to test whether models can refrain from taking unsafe actions when no safe operational alternatives are available. The dataset includes two subsets: Safe and LT (Life-Threatening), which correspond to different clinical risk levels. It supports both textual and multimodal inputs, and provides two abstention settings: Trivial Abstention (hide the question and only display options) and Standard Abstention (display the complete question and options). Additionally, prompt variants tailored for reasoning and non-reasoning models are provided. The dataset fields include ID, source dataset, question text, options, answer index, image, modality type, original options, original answer, and multiple prompt fields.




