10 records · Browse by task

Healthcare benchmarks

Find tests for health conversations, clinical documentation, billing, and medical questions.

Conversation rubrics, medical exams, and administrative tasks measure different work. Their scores do not establish clinical safety or patient outcomes.

Records keep metrics and harness variants separate. Source rows can include different model and agent configurations; linked profiles count distinct public model identities. Neither count represents independent evaluations. Dates describe the stored source record.

Public model rows count finite, displayable results linked to reviewed original artifacts. Publisher comparison charts can include other providers' models. Follow the result link to inspect each row's source and run conditions.

Health and clinician conversations

HealthBench (raw)
7 public model rows linked to original artifacts
Original source
The raw HealthBench score stays separate from the length-adjusted score.
5,000 multi-turn patient conversations
HealthBench (length-adjusted)
7 public model rows linked to original artifacts
Original source
A length adjustment changes the metric; inspect it beside the raw score.
5,000 multi-turn patient conversations
HealthBench Hard
12 public model rows linked to original artifacts
Original source
The harder subset is a separate evaluation. Inspect row notes for raw or length-adjusted conditions.
1,000 health prompts
HealthBench Professional
13 public model rows linked to original artifacts
Original source
Clinician-facing tasks can include product harnesses; read the run conditions.
Clinician chat tasks
HealthBench Professional (raw)
7 public model rows linked to original artifacts
Original source
The raw Professional score stays separate from the length-adjusted result.
525 physician-authored conversations

Documentation and billing

MedScribe
109 stored source rows · 80 linked model profiles
Source update in snapshot: 2026-10-13
Source snapshot captured: 2026-10-15T18:47:28.403Z
Original source
SOAP-note documentation is evaluated from transcripts; it does not measure an ambient-scribing product end to end.
Medical administrative support tasks
MedCode
107 stored source rows · 78 linked model profiles
Source update in snapshot: 2026-10-13
Source snapshot captured: 2026-10-15T18:47:28.403Z
Original source
Medical billing support is administrative evidence, separate from diagnosis or treatment.
Medical billing support tasks

Medical questions

MedXpertQA (Text)
5 public model rows linked to original artifacts
Original source
Text-only medical multiple-choice results remain separate from the multimodal test.
2,450 medical multiple-choice questions
MedXpertQA (MM)
8 public model rows linked to original artifacts
Original source
Medical-image multiple-choice results use their own task set.
2,000 multimodal medical questions

Historical evidence

MedQA
95 stored source rows · 39 linked model profiles
Archived source board
Source update in snapshot: 2026-04-16
Source snapshot captured: 2026-10-15T18:47:28.403Z
Original source
Vals archived this medical-question board after saturation; it is not a current-release test.
Medical question answering

Other collections

All benchmarks