MedExpert

A clinician-annotated benchmark for evaluating the factuality and completeness of medical chatbots.

MedExpert is an expert-annotated dataset for evaluating medical chatbots on two safety-critical axes: factuality and completeness (omission detection).

  • 540+ clinician-annotated question–response pairs from high-risk specialties (prenatal care and young-adult mental health), plus 32 dual-annotated pairs.
  • 100+ unique questions authored by clinicians, each answered by 5 open-source LLMs (Llama-2 7B, Llama-3.3 70B, OLMo-2 13B, Gemma-2 27B, OpenBioLLM-70B).
  • Annotated by 8 practicing clinicians (MDs, residents, LCSW), with severity ratings for every factuality error and omission.

Paper · GitHub · Dataset. Presented at ML4H 2025.

(Yarmohammadi et al., 2025)

References

2025

  1. MedExpert: An Expert-Annotated Dataset for Medical Chatbot Evaluation
    Mahsa Yarmohammadi, Alexandra DeLucia, Lillian C Chen, and 8 more authors
    In Machine Learning for Health 2025, 2025