ONNX weights
Hey, here too: could you add the onnx weights? Thanks!
Hey @do-me , the conversion code I added to the repo unfortunately doesn't work with this model since this was still using our word_level tokenizer, so I'm not sure if this is easy or possible to do. We've moved away entirely from using that (hence the M2V_base_glove_subword model, which we created to replace M2V_base_glove). However, both these models get outperformed by the much smaller potion-base-8M model on all tasks, so it's probably better to just support those models instead.
I was already wondering whether these models are now kind of obsolete. Maybe, do you think it's a good idea to add a note that folks just discovering these models are guided towards the best (right now Potion-family)?
So in a nutshell atm, one can sum it up like this, right? This would be the model selection I allow in a downstream app.
That's a great idea actually! The old models are indeed obsolete, I will add a note to the model cards. We did not want to remove them since that would break current integrations. As for your overview: yes, this is correct (I will add weights for the science model as well btw, I'm still experimenting with that one a bit and will most likely also release a slightly larger version of that soon). For most users, potion-base-8M is the model we recommend. I would also exclude M2V_base_output since that one is essentially replaced by potion-base-8M (same size, but better).
Perfect, you can use all of these models in the app now if you like, will announce it these days :)
https://github.com/do-me/semantic-similarity-table
Awesome, super cool to see! Looking forward to the announcement :)