DEJIMA: 3.88M Japanese Image-Caption and Image-QA Pairs

DEJIMA is a large-scale Japanese multimodal dataset containing 3.88 million image-caption pairs and 3.88 million image-question-answer pairs. It was created by MIL-UT using a reproducible pipeline involving web-scale image collection, strict filtering, evidence extraction, and LLM-based annotation under grounding constraints. The dataset was last updated on December 2, 2025.

Use Cases

• Training Japanese image captioning models based on the 3.88M image-caption pairs.
• Fine-tuning visual question answering models for Japanese based on the 3.88M image-QA pairs.
• Benchmarking multimodal models on Japanese-language grounding and evidence extraction tasks.
• Developing or evaluating scalable data generation pipelines for non-English multimodal datasets.

Strengths

• Contains 3.88 million image-caption pairs (DEJIMA-Cap).
• Contains 3.88 million image-question-answer pairs (DEJIMA-VQA).
• Annotations are generated under grounding constraints, which may improve quality.
• Constructed through a scalable and fully reproducible pipeline.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect bias inherent to web-scale image collection.

Provenance

Source

MIL-UT

Collection Method

Web-scale image collection, strict filtering and deduplication, detection-driven evidence extraction, LLM-based caption/VQA generation.

Freshness

Last updated 2025-12-02 03:27:56; freshness should be verified.