FMBench
收藏资源简介:
FMBench是由布里斯托大学、帝国理工学院、伊利诺伊理工学院和慕尼黑工业大学联合创建的医学多模态数据集,包含30,000个医学视觉问答对和10,000个医学图像报告对。数据集详细标注了种族、性别、语言和民族等多样性属性,旨在全面评估多模态大语言模型在医学任务中的公平性。数据集的创建过程结合了哈佛-FairVLMed数据集,通过LLM生成高质量的问答对,并进行后处理优化。FMBench的应用领域主要集中在医学视觉问答和报告生成任务,旨在解决模型在不同人口群体中的公平性问题。
FMBench is a medical multimodal dataset jointly developed by the University of Bristol, Imperial College London, Illinois Institute of Technology, and Technical University of Munich. It encompasses 30,000 medical visual question-answer pairs and 10,000 medical image-report pairs. This dataset is comprehensively annotated with diversity attributes including race, gender, language, and ethnicity, aiming to conduct comprehensive evaluations of the fairness of multimodal large language models (LLMs) in medical tasks. The construction of FMBench integrates the Harvard-FairVLMed dataset, where high-quality question-answer pairs are generated via LLMs and optimized through post-processing steps. The primary application domains of FMBench focus on medical visual question answering and report generation tasks, with the objective of addressing fairness issues of models across diverse demographic groups.




