M3-MedQA: Multilingual Medical VQA Benchmark in Five Languages

M3-MedQA contains between 1,000 and 10,000 medical image-question pairs across five languages, developed by pnu-clink in 2024. It extends the WorldMedQA-V dataset to evaluate cross-lingual consistency and medical reasoning in English, Korean, Japanese, Arabic, and Wolof.

Use Cases

• Evaluating cross-lingual consistency by comparing model outputs for the same image across the five language versions of the 'questions' field
• Benchmarking medical reasoning using the 'options' field for multiple-choice selection
• Training multimodal models to associate medical imagery with localized clinical terminology

Strengths

• Multilingual support for five languages including the low-resource language Wolof
• 1,000 to 10,000 record scale
• MIT licensed for open redistribution

Limitations

• Relatively small sample size of under 10,000 records
• Potential for translation-induced noise in non-English subsets

Provenance

Source

pnu-clink (extension of WorldMedQA-V)

Freshness

Last updated March 2026.

Geography

Global