Physician-graded clinical rubrics paired with Stanford MedHELM hospital tasks
Seventeen protocols feed the healthcare vertical, led by HealthBench Professional (19%), HealthBench Hard (10%), HealthBench Professional (raw) (9%) alongside MedXpertQA (text and multimodal), PhysicianBench, FrontierBioML, and eight Apache-2.0 Stanford CRFM MedHELM tracks covering clinical decision support, SOAP note generation, MedCalc-Bench bedside equations, MedHallu hallucination resistance, PubMedQA, and administrative workflows.
Combining multi-turn physician conversation rubrics with hospital paperwork and equation calculators prevents models memorized on multiple-choice USMLE flashcards from dominating practical clinical-documentation rankings.