Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• Jagle is the largest Japanese multimodal post-training dataset to date, comprising approximately 9.2 million instances.result 95%
• Jagle was constructed using heterogeneous source data including images, image-text pairs, and PDF documents rather than relying on existing VQA datasets.method 90%
• VQA pairs in Jagle were generated through multiple strategies including VLM-based QA generation, translation, and text rendering.method 90%
• A 2.2B model trained with Jagle surpasses InternVL3.5-2B in average score across ten Japanese evaluation tasks.result 90%
• The Jagle-trained 2.2B model approaches within five points of Qwen3-VL-2B-Instruct on Japanese evaluation tasks.result 90%
• Combining Jagle with FineVision does not degrade English performance and instead improves it compared to training with FineVision alone.result 90%
• Existing VQA datasets in non-English languages remain limited in both scale and domain coverage, posing a major obstacle to building high-quality multilingual VLMs.limitation 85%
• The dataset, trained models, and code for Jagle are being released to facilitate reproducibility and future research.technical 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card