Will the component models be available separately?

#31
by libryo-ai - opened

M3 is great, thanks for sharing your work. Will the dense and multi-vector parts be available separately? eg bge-large-m3 and bge-colbert-m3? I would like to try optimise ONNX for deployment at scale.

Beijing Academy of Artificial Intelligence org

Thanks for your interest in our work!

You can use transformers or sentence-transformers to load BAAI/bge-m3 following: https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/baai_general_embedding#using-huggingface-transformers. In this way, it will only load the dense model.

For other retrieval modes, I find some resource from community:https://huggingface.co/aapot/bge-m3-onnx. https://huggingface.co/hooman650/bge-m3-onnx-o4

Sign up or log in to comment