Abliteration Process
#3
by
thedarktrumpet
- opened
Good morning,
I've been trying out abliteration, trying to understand it a bit better and the different methods of going about it. I've been trying a newer method https://github.com/jim-plus/llm-abliteration, but have had some really weird results.
Do you mind sharing the commands and/or config you're using? In my case, I tried targeting layers 42-47. The strength/scale matters a lot it seems. Mostly doing it for learning reasons, but if you don't mind sharing more details on your process, I'd appreciate it greatly.
The choice we made is for 47 layers.