Skip to content

Conversation

@matthewhaynesonline
Copy link
Contributor

Add implementation for https://huggingface.co/naver/provence-reranker-debertav3-v1. This is still a WIP, but I wanted to gauge interest before going too far.

Notes

  • Provence has a CC Non Commercial license
    • Not sure if that's against candle conventions to implement models with restricted licenses
  • Since it's based on DeBERTa-v3, the core implementation is pretty easy (forward) however the original py transformers implementation provides a custom process function (and thus requires remote code)
    • Since by convention the candle models implement forward but not full pipelines, I wasn't sure if adding a process function in the model would be a bad choice.
    • For now it's there, but feature flagged
  • I started by looking at the DeBERTa example and the XLM-RoBERTa reranking and tried to go from there.
  • There's other versions (multilingual) and an LLM approach, but this particular model was a good fit for a use case of mine, so that's why I've focused on it
  • If there's potential for approving this, I can reach out to the original authors on HF to get their feedback as well

Thank you!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant