Open cdoern opened 3 months ago
Note that with: https://github.com/instructlab/instructlab/pull/2065 eval will no longer support hf paths. Currently it is only supported for mmlu.
We actually could support hf downloads but the current implementation is insufficient because it would rely on secondary libraries doing the model downloads which means the downloads don't end up in our cache. To support we would need to integrate ilab download capability into eval.
--model
to --model-path
imo makes much more sense to use the same convention for local paths across local libraries, and then maybe --model-repo
as you suggested for HF
@alinaryan @jaideepr97 - I compiled all the issues where this is mentioned in into the main description. do you want to make this the main issue/epic?
the clickext class should be extended to figure out if a flag is either an HF path OR a local path. Based off of which type of argument is passed, there should be a log and specific behavior for the different types.
in
ilab model train
here are the current issuesBut ALL of them use
--model-path
adding some logic to the custom option class could resolve this and download the model when necessary.in
ilab model evaluate
--model
can be either a HF repo OR a local path.expanding the clickext class to handle all of these will allow us to handle these scenarios in one place rather in the various cmds while still maintaining the single flag method.
Alternatively
all of these flags can be split into
--model-repo
and--model-path
flags (one for HF/OCI and one for local paths)Issues resolved when this issue closes: