Diabetes Question-Answering Model Evaluation Test Set

This test set was curated through a systematic selection process based on the CMtMedQA dataset (huggingface.co/datasets/Suprit/CMtMedQA). CMtMedQA, provided by Suprit, contains approximately 70,000 authentic multi-turn doctor–patient dialogues across 14 clinical departments and is widely used as a foundational resource in Chinese medical dialogue research.To construct the present test set, candidate samples were first retrieved from the original CMtMedQA dialogues using domain-specific keywords such as “diabetes,” “endocrinology,” and “blood glucose,” yielding a subset relevant to chronic disease scenarios.

Subsequently, in accordance with the requirements of multi-turn interaction, only dialogues in which the patient contributed more than three turns were retained. All candidate dialogues underwent case-by-case verification by clinical experts, who reviewed the medical topics, content completeness, and linguistic accuracy. After this expert validation process, a total of 100 authentic multi-turn patient dialogues were selected to form the released test set, providing data support for question-answering tasks related to chronic disease management.