In response to the growing need for high-quality, domain-specific evaluation resources for French medical NLP, we introduce DiabetesQA, a novel question–answer dataset dedicated exclusively to diabetes. The dataset is entirely authored in French and comprises 41,992 carefully curated question–answer pairs. DiabetesQA is designed to reflect both clinical and patient-oriented perspectives by integrating content from reliable French-language sources, including medical textbooks, peer-reviewed patient forums, and expert-validated educational materials. Question–answer pairs were automatically extracted from the source documents and subsequently reviewed by bilingual medical specialists to ensure medical accuracy, clarity, and internal consistency. To support robust and reproducible model evaluation, the dataset is partitioned into training (33,593 instances), validation (4,199 instances), and test (4,200 instances) sets.