Abliteration but still not entirely uncensored
#6
by
augustine-aisg
- opened
As shown in the image below
Try introducing it using other prompt words.
Try introducing it using other prompt words.
Yea, may work. But ive noticed this with lots of your models. What do you think is causing the censorship to remain?
Would be interested to know how this could be improved.
It could be due to probability, simple guidance, or it may affect attention.
It could be due to probability, simple guidance, or it may affect attention.
Hi, thanks for getting back to me so quick.
I was just wondering.
Do you have future plans to improve the process?
Are there any research papers or resources which you recommend. To further improve this process?
It may need a brand new uncensored data source to entirely remove all these censorship.