DiagTrace: Traces of Diagnostic Reasoning in Clinical Cases

DiagTrace is a bilingual (French/English) dataset for evaluating the diagnostic reasoning capabilities and final diagnosis accuracy on real-world clinical cases. It contains 180 clinical cases spanning multiple medical specialties, paired with model predictions from 5 LLMs and expert scores from two medical doctors across 6 evaluation criteria. The dataset organized into two subsets: "cases": Clinical cases with ground truth "predictions": LLM predictions with expert scores (One row per (case × model × language) combination — 180 cases × 5 models × 2 languages = 1,800 rows)