JDocQA Reasoning Bench: A Japanese Document Question-Answering Benchmark

Ricoh AI developed the JDocQA Reasoning Bench to evaluate the reasoning performance of multimodal large language models on documents containing figures and charts. The benchmark was created as part of Japan's GENIAC project, which aims to strengthen domestic generative AI development capabilities. It was published on the Hugging Face platform by the author ricoh-ai.

Use Cases

• Benchmarking LMM reasoning performance based on the described focus on documents with figures and charts.
• Training models for Japanese document question-answering based on the dataset's stated purpose.
• Evaluating multimodal understanding of text and visual elements within documents.

Strengths

• Dataset was developed by a corporate research team (Ricoh AI) for a formal government-backed project (GENIAC).
• Specifically designed to measure reasoning performance on a challenging multimodal task (documents with figures/charts).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Sample data is unavailable, preventing preview of data structure and content.

Provenance

Source

Ricoh AI, developed under Japan's GENIAC (Generative AI Accelerator Challenge) project.

Collection Method

Created for model evaluation; specific data gathering method is not detailed.

Freshness

Last updated 2026-05-19 11:12:20; freshness should be verified.

Geography

Japan (inferred from the project's domestic focus and Japanese language content).