←back to thread

Uncensor any LLM with abliteration

(huggingface.co)
586 points mizzao | 2 comments | 13 Jun 24 03:42 UTC | HN request time: 0.434s | source
1. seydor ◴[13 Jun 24 05:45 UTC] No.40666318[source]▶
The word is 'ablation'. Do not butcher it
replies(1): >>40666485 #
2. girvo ◴[13 Jun 24 06:17 UTC] No.40666485[source]▶
Amusingly, they explicitly call it this in the article itself:

> Once we have identified the refusal direction, we can "ablate" it, effectively removing the model's ability to represent this feature