Jagle addresses a persistent blind spot in multimodal model training: large-scale, high-coverage Japanese post-training data. By assembling roughly 9.2 million instances across images, PDF-derived documents, and synthetic VQA, Jagle supplies the language- and script-specific signal most VLMs lack — and the maintainers report measurable gains on Japanese tasks for models trained with the corpus.
What Sets It Apart
- Japan-focused scale: ~9.2M multimodal examples specifically curated or translated for Japanese, rather than relying on sparse Japanese subsets of global datasets — this matters because script, cultural context, and annotation styles differ from English-centric corpora.
- Multiform input types: combines image–text pairs (e.g., WAON, japanese-photos), PDF corpora (FinePDFs-Edu, NDL Warp PDFs) and chart/table datasets (PlotQA, TAT-QA), increasing coverage of real-world VQA and document understanding scenarios.
- Hybrid VQA generation pipeline: leverages VLM-based QA generation plus translation and filtering to expand question–answer pairs in Japanese, enabling more robust VQA supervision without entirely manual annotation.
- Empirical impact: training runs reported on a 2.2B model (Qwen3-1.7B + SigLIP2 vision encoder) show notable improvements on Japanese benchmarks after post-training with Jagle.
Who it's for — and tradeoffs
Great fit if you are training or fine-tuning vision–language models that must handle Japanese text, document images, or culturally specific visual context; research teams needing large-scale Japanese VQA and document-reading signals will find the corpus useful. Look elsewhere if you require fully redistributable, globally mirrored data (Jagle is hosted domestically and governed by Article 30-4 of the Japanese Copyright Act), or if your project mandates permissive open licenses for all derived content. Also expect heterogeneous source licenses and attribution requirements because Jagle aggregates many upstream datasets.
Where it fits
Use Jagle as a post-training / continued-pretraining corpus to close the Japanese coverage gap before instruction tuning or downstream fine-tuning for VLM tasks. It complements English-centric multimodal datasets rather than replacing them, and was used to train llm-jp-4-VL models as an example downstream application.
