Are you working on uncensoring 405B and lorablating it?
#6
by
User8213
- opened
I am just asking so a uncensored model could finally be out there.
Lorablating (also called abliteration from adapter) works by extracting the changes abliteration makes to a model and applying it to another model, e.g. taking the difference between llama 8b 3 and llama 8b 3 abliterated, then applying it to llama 8b 3.1
There's no abliterated 405b, since there was no llama 3 405b, so there's no way to make a LoRA from it. Abliterating it in general for 405b is a bit hard because the 3.1 family seems to have more safety aligned layers than 3