Benchmark RadarJMed48k
JMed48k evaluates vision-language models on 48,862 Japanese medical licensing exam questions from 11 national examinations (2005-2025), with images annotated under an 8-type taxonomy. The JMed48k-Eval subset contains 12,484 scored questions, including text-only and with-image items, scored separately.
- Released
- 2026-05-21
- Readiness
- Paper only
- Primary field
- Health & Life Sciences
Why it matters
Provides a profession-stratified evaluation for vision-language models in Japanese medical licensing, enabling comparison of image use across professions and model types, and addressing the lack of multilingual medical benchmarks.
Motivation
We introduce JMed48k, a multi-profession Japanese healthcare licensing benchmark for evaluating vision-language models.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.