←back to thread

Uncensor any LLM with abliteration

(huggingface.co)
586 points mizzao | 2 comments | 13 Jun 24 03:42 UTC | HN request time: 0.458s | source
Show context
Mathnerd314 ◴[13 Jun 24 05:28 UTC] No.40666249[source]▶
Reminds me of https://vgel.me/posts/representation-engineering/. There they were adding a control vector, w' = cvec + w, here they are "ablating" it, w' = w - dot(w,cvec)*cvec. There is an interesting field of learning how to "brain chip" LLMs into doing what you want.
replies(1): >>40666307 #
1. Der_Einzige ◴[13 Jun 24 05:43 UTC] No.40666307[source]▶
There's so much work just like this coming out simultaneously.

Steering Vectors, Control Vectors, PyReft, PeFT improvements, Obliteration. It's a great time to be doing representation engineering.

replies(1): >>40670760 #
2. Mathnerd314 ◴[13 Jun 24 15:19 UTC] No.40670760[source]▶
There is some difference between fine-tuning with PyReft / PeFT, the approaches here are more on-the-fly. Like you can regenerate the control vectors from prompts in a few seconds.