search

allenai / open-instruct

Apache License 2.0
1.1k stars 145 forks source link

Trust HF OLMo models. #111

Closed dwadden closed 5 months ago

dwadden commented 5 months ago

Previously, attempting to load OLMo models in a Beaker job raised an error because these models run custom code. This fixes it by setting trust_remote_code=True for OLMo models.

At some point we should just unify the command line parsers for the different eval tasks and allow the user to pass --trust_remote_code as a command line flag; this will take a bit of a refactor.

dwadden commented 5 months ago

Hmm I see. Does adding this dependency actually break things for other HF models? The reason to add this would be: if we wanna use open-instruct internally to run evals of OLMo models, then it makes sense to be able to run OLMo models by default. But if this breaks other things, agreed this doesn't work.

hamishivi commented 5 months ago

I don't think it does, but I don't want extra packages that might mess with someone's environment that aren't actually required...

dwadden commented 5 months ago

Got it. I removed it from requirements.txt and added a note in the readme. I also added it to the Dockerfile; lmk if you disagree with this and I can remove from there as well.

hamishivi commented 5 months ago

Dockerfile, is fine, lgtm!