Should large still exist? Or should it link to large-v2?
I think open_ai released large-v2 and now if you don't specify manually large-v1 then you get v2 by default.
Should this be the case here?
For backward compatibility, we are leaving this as the v1. It makes more sense IMO
Indeed, we're following a backwards compatibility convention here:
openai/large-> large V1 modelopenai/large-v2-> large V2 model
Hey
@sanchit-gandhi
What I was suggesting is to align this with openai/whisper :
See below, they are linking large to large-v2 for every new download, and if someone really wants the older v1, they will need to ask for it directly.
I'm ok either way, just wanted to suggest to keep in alignment so folks won't get confused.
I see! Thanks for clarifying
@altryne
! I think here we probably want to keep large and large-v2 distinct for backwards compatibility (something we value very strongly in transformers! No-one wants code that changes unexpectedly after a remote PR change!)
We can add a message to the README of the large model stating that the large-v2 checkpoint is preferred?
Edit: opened a PR here https://huggingface.co/openai/whisper-large/discussions/24