Thank you for M3

[jobergum]

Thank you for sharing the M3 model with the community! This is a welcome direction, where you can get multiple representations from the same model!

One comment and two questions

For sparse retrieval methods, most open-source libraries currently do not support direct utilization of the BGE-M3 model. Contributions from the community are welcome.

Vespa allows supporting all three different representations, I created this notebook to demonstrate how to use it with Vespa.

I noticed that the model hub repo isn't tagged with a license like the previous models; I assume this is an oversight? Secondly, did you consider training with a lower dimensionality for the linear colbert layer? The original colbert paper has good results with a reducing the dimensions. I think it would be more useful that way and I suspect that it would not degrade accuracy much.

Again, thank you!