search

kevinyaobytedance / llm_unlearn

LLM Unlearning
MIT License
95 stars 12 forks source link

Doubt the pratical perfermance of LLM Unlearning #5

Open shenjiyuan123 opened 2 months ago

shenjiyuan123 commented 2 months ago

Hi, thanks for your work. I also agree with the value of LLM unlearning in privacy-preserving since I also work on the unlearning of Federated Learning scenario. However, I just tested a simple case and used the original facebook/opt-1.3b model as follows:

Capture Capture

The output is already whitespace, which makes me doubt the true effect of the proposed method.

Hope to hear further explanation from you. Thanks!

dashfahojg commented 1 month ago

Oh，I also don't think this code is effective. maybe you can try llama2-7b?