JaWildText is a Japanese scene text understanding benchmark designed to evaluate vision-language models on text-rich real-world images. It is authored by llm-jp and was last updated on May 25, 2026. The benchmark focuses on diagnosing OCR, visual question answering, and structured information extraction in practical conditions.
Use Cases
- Evaluate Japanese OCR performance based on dense signboards and handwritten text mentioned in the description
- Benchmark scene text visual question answering capabilities for VLMs based on text-rich images
- Test structured information extraction models based on mobile-captured receipts
Strengths
- Benchmark is designed for practical conditions including dense signboards, handwritten text, and mobile-captured receipts
- Last updated on May 25, 2026
Limitations
- Column-level documentation is absent; field semantics must be inferred after download
- Row count is unknown, which may limit suitability assessment
Provenance
- Source
- llm-jp
- Freshness
- Last updated 2026-05-25 00:49:55