Does Candle have an implementation of the classification head for BERT or should I implement it myself?

[manifest]

I've got model parameters in safetensor format by fine-tuning a BERT model with the classification head using Transformers library (i.e. using AutoModelForSequenceClassification model class). Now, I want to perform inference using Cande library.

Does Candle have an implementation of the classification head for BERT or should I implement it myself?

[manifest]

I've implemented a classification head for DistilBertModel similarly to DistilBertForSequenceClassification implementation of Transformers library. Does it make sense to contribute the code to candle_transformers crate?

[Narsil]

You can definitely share the example, it'll help others for sure.

As for the inclusion within candle-transformers, I'm not sure we have taken a stance on implementing all the potential heads in transformers. So we might not merge it or not soon. Currently bert is mostly used (to the best of my knowledge) for RAG implementations, which is why it has the current form. Adding the classification in client code seems the most reasonable for me at the present.

[gilljon]

Mind sharing the implementation? Trying to understand how to use BERT for ranking :)