edit is successful but prompts stay the same

[nrosen1]

tested with gptj + llama-13 models from huggingface

[arnab-api]

Hi,

I need a little more context to look into this issue. How are you calculating your LREs? Does it seem to work for an example?

And, from the code snippet you provided I think the problem is how you are calling the editor. mt.model inside the editor is just a reference to the model itself. It will work the same. So,

First you should approximate an LRE (checkout this or the demo. Then, create the editor object like this,

editor = editors.LowRankPInvEditor(
    lre=lre,           # the LRE operator approximated on a set of samples
    rank=rank,    # rank of the edit
)

And, then you call the edit

result = editor(
     subject = sample.subject,
     target = target.subject,
)

If you think this is not the problem and you are doing everything properly, please provide the full script to reproduce this issue. And, I will take a look.