Visual Document Retrieval - New vdr-2b-multi-v1 Multilingual Model Released

AI Impact Summary

The team has released vdr-2b-multi-v1, a new multilingual embedding model designed for visual document retrieval, built upon MrLight/dse-qwen2-2b-mrl-v1 and trained on a 500k synthetic multilingual dataset spanning six languages (Italian, Spanish, English, French, German). This model addresses the scarcity of high-quality multimodal datasets, offering improved cross-lingual retrieval and faster inference compared to previous models, particularly through techniques like Matryoshka representation learning and optimized VRAM usage.