Update similarity_fn_name = 'cosine' to MaxSim

[sam-hey]

When loading a model, the similarity_fn_name attribute defaults to cosine instead of max_sim. Since SentenceTransformer does not support max_sim natively, this functionality must be implemented using pylate.

self.model = colbert_model.ColBERT("colbert-ir/colbertv2.0")

[NohTow]

Hello,
Yes indeed, I did not overwrite it as we are not directly using it in the training/inference code.
It happens that I am working on another PR to add NanoBEIREvaluator and I added the SimilarityFunctions to PyLate, with only MaxSim for now.
See #76 for when it gets merged.
Also, at some point I should change the code to use these calls instead of directly calling scores in the training and inference, but that will be another PR.

[sam-hey]

Awesome, thanks for the quick response—I'm really looking forward to it!

By the way, I absolutely love this project—fantastic work!

[NohTow]

Thanks for the kind words!
Actually, could you try the branch with your use case (I suppose you have one) and report if anything is not working for you?
Maybe it will be in another PR, but it's nice to see if there is something wrong that I could fix.
The PR itself should be merged soon, but if you want to try it now, it might prevent merging a bug!

[sam-hey]

Sure, it works for me without any issues!
Thanks for the implementation—I’m looking forward to the new version being available!