HakushoBench: Japanese Chart and Table VQA Benchmark from Government White Papers

HakushoBench is a Japanese visual question answering benchmark built from 33 governmental white papers. It contains 2,053 images spanning over 10 chart and table types, with manually annotated QA pairs. The dataset was created by llm-jp and last updated on Hugging Face in June 2026.

Use Cases

• Benchmarking multimodal models on Japanese chart comprehension based on the annotated QA pairs.
• Training models for holistic understanding of complex government data visualizations.
• Evaluating AI performance on diverse image types from official documents.

Strengths

• Contains 2,053 images from a specific, authoritative source (33 governmental white papers).
• Includes manually annotated QA pairs designed to assess deep understanding.
• Covers over 10 different chart and table image types.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect the geographic and institutional bias inherent to Japanese government white papers.

Provenance

Source

llm-jp via Hugging Face.

Collection Method

Built and manually annotated from 33 Japanese governmental white papers.

Freshness

Last updated 2026-06-02 02:48:52; freshness should be verified.

Geography

Japan (inferred from source documents).