Unrecognized 'num-gen-tokens' Argument in model_args.json During Model Checkpoint Pruning about gill HOT 4 CLOSED

Hi, thanks for pointing this out! This is a typo and it should be num_tokens -- I think this was an artifact left from early in development. I believe 8216fea should fix this, but please let me know if not.

from gill.

That's what I was thinking too :)

Also, loading a new trained model, for training the decision classifier fails in the "TrainDecisionClassifier" fails too. It is due to the assertion error on line 45 of models.py, saying:
AttributeError: 'args' object has no attribute 'text_emb_layers'

There are 3 arguments that are not saved in the model_args.json file (compared to provided json file in the repo) after conducting a training:

"text_emb_layers": [
-1
],
"share_ret_gen": true,
"norm_image_embed": "none"

I think this is the reason.

P.s: I am just trying to train the model with a different LLM and rerunning all the scripts

from gill.

Thanks! This is helpful. The other arguments are not used, but text_emb_layers is, so we need it to be saved in the model_args.json file. Oddly enough it doesn't seem to be saved here, even though it's part of model_args. I guess we can explicitly add it, since there's no command line flag to change text_emb_layers in main.py: 6b183ac

Sorry about that, text_emb_layers was something used for debugging and I didn't remove it completely in the final version. I think with your newly trained LLM, you can also edit the .json to set "text_emb_layers": [-1] and it should work fine. Hope that helps!

from gill.

Of course :) adding "text_emb_layers": [-1] explicitly worked actually, thanks!

I see that the "share_ret_gen" is used here , but I guess it's not affecting anything.

thanks anyways.

from gill.