Hi, I am interested with the multi-hop problem and would like to clarify whether CaKE is evaluated in lifelong editing or only one-time editing?
The C.4 appendix shows multiple edit test are evaluated. I assume the edit_num is either the batch size or the number of edits. However, the code indicates the LoRA adapters are unloaded after every edit. So, is it correct to assume edit_num refers to the batch size?
|
model = model.unload() |
|
del model.peft_config |
Thank you
Hi, I am interested with the multi-hop problem and would like to clarify whether CaKE is evaluated in lifelong editing or only one-time editing?
The
C.4appendix shows multiple edit test are evaluated. I assume theedit_numis either the batch size or the number of edits. However, the code indicates the LoRA adapters are unloaded after every edit. So, is it correct to assumeedit_numrefers to the batch size?CaKE/edit_utils.py
Lines 321 to 322 in df528e1
Thank you