DrugAudit
收藏资源简介:
DrugAudit是由佛罗里达大学和澳门大学联合创建的一个权威感知型药物信息问答基准数据集,包含3,772个精心设计的评估项目。该数据集覆盖了九个源数据库子集,并整合了MedQA和PubMedQA中药物相关的子集,旨在通过双评委大语言模型协议对答案的上游黄金源匹配度、令牌级语义片段重叠和引用忠实性进行多维度评分。其构建过程严格遵循证据等级区分原则,特别强调区分原始监管记录与下游聚合器来源,以解决现有生物医学问答评估中缺乏权威性溯源和引用质量量化的问题。该数据集主要应用于药物信息检索与问答系统的性能评估,致力于提升临床决策支持系统中答案的可验证性与溯源可靠性,从而保障患者安全。
DrugAudit is an authority-aware drug information question answering benchmark dataset jointly developed by the University of Florida and the University of Macau, which comprises 3,772 meticulously designed evaluation items. It encompasses nine source database subsets, and incorporates drug-related subsets extracted from MedQA and PubMedQA. Its core objective is to conduct multi-dimensional scoring of answers across three key aspects: the matching degree between answers and their upstream gold-standard sources, token-level semantic fragment overlap, and citation faithfulness, via the dual-judge large language model protocol. Its construction strictly adheres to the principle of evidence hierarchy, with particular emphasis on differentiating between original regulatory records and downstream aggregator sources, aiming to address the gap that existing biomedical question answering evaluations lack authoritative traceability and quantifiable citation quality. This dataset is primarily applied for performance evaluation of drug information retrieval and question answering systems, and is committed to improving the verifiability and traceable reliability of answers in clinical decision support systems, thereby safeguarding patient safety.




