>>105741413
>By "inhibiting certain behaviors and concepts in the trained model", they almost certainly means that they taught the model that certain concepts need to return the same image.
yeah I agree with you, they didn't specifically target layers but by doing this kind of training, there's in consequences some special layers that activate to cuck the output, that's what's happening on the llm models, and for them they manage to locate those layers and they can remove them
https://huggingface.co/blog/mlabonne/abliteration