lion-ai/MedKG-QA
收藏资源简介:
MedKG-QA是一个波兰语医学问答数据集,包含5086个经过验证的医学知识控制问题。这些问题从医学知识图谱MedKG生成,并通过自动质量门过滤,确保高质量。数据集组成包括跳数分布:1跳844个、2跳653个、3跳165个、4跳166个、5跳3258个;其中事实(基于1-2跳关系)1497个,案例(基于3+跳关系,模拟临床病例)3589个。平均质量评分(由LLM法官评估,范围1-5)为4.94。验证过程包括四个步骤:答案与源证据的一致性、基于源片段的可解性、无答案泄露、以及由独立LLM法官评估质量(评分≥4/5)。数据列包括id、问题(波兰语)、答案、答案别名、跳数、类型(事实或案例)、质量评分、推理过程和证据(来自知识图谱的事实和引用)。数据来源包括波兰维基百科(CC BY-SA 4.0)、药物说明书(来自URPL注册,CC BY 4.0)和ICD-10波兰版。数据集基于CC BY-SA 4.0许可证发布。注意:问题由LLM生成并自动验证,可能包含错误,不构成医疗建议。
MedKG-QA is a Polish-language medical knowledge-grounded question answering dataset containing 5,086 validated questions. These questions are generated from the medical knowledge graph MedKG and filtered via an automatic quality gate to ensure high quality. The hop count distribution of the dataset is as follows: 844 1-hop questions, 653 2-hop questions, 165 3-hop questions, 166 4-hop questions, and 3,258 5-hop questions. Among them, there are 1,497 factual questions (based on 1-2 hop relations) and 3,589 case questions (based on 3+ hop relations, which simulate clinical cases). The average quality score, evaluated by LLM judges with a rating range of 1 to 5, is 4.94. The validation process includes four steps: consistency between the answer and source evidence, solvability based on source snippets, absence of answer leakage, and quality assessment by independent LLM judges requiring a score of ≥4/5. The dataset columns include id, question (Polish), answer, answer aliases, hop count, type (factual or case), quality score, reasoning process, and evidence (facts and citations from the knowledge graph). Data sources include Polish Wikipedia (CC BY-SA 4.0), drug package inserts (from the URPL registry, CC BY 4.0), and the Polish edition of ICD-10. The dataset is released under the CC BY-SA 4.0 license. Note: Questions are generated by LLMs and automatically verified, may contain errors, and do not constitute medical advice.




