Tests whether a model can compute clinical values (dosages, risk scores, dates) from a patient note the way a bedside medical calculator would, graded by exact match or numeric tolerance.
unassessed
| Category | domain |
|---|---|
| Subcategory | clinical/medical calculation |
| Page status | active |
| Metric | accuracy (exact match for rule-based and date calculators; within 5% tolerance for equation-based lab/physical/dosage calculators) |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 11643 |
| Dataset licence | CC-BY-SA-4.0 |
MedCalc-Bench gives a model a patient note and a question naming a specific clinical calculation — a creatinine clearance, a Glasgow Coma Score, a gestational age, a drug dosage — and asks it to extract the relevant values from the note, apply the correct formula or clinical rule, and produce the answer a bedside calculator would give. This targets a narrower, more mechanical skill than open-ended medical question answering: correctly identifying which numbers in a note matter, and computing with them correctly, rather than recalling medical facts in prose.
Given a patient note (drawn from case reports or clinical vignettes) and a question naming one of 55 calculators, the model must extract relevant entities and output a numeric or categorical answer with its computation shown.
No model card in ModelSpec reports this benchmark yet.