Comment by nottorp
12 days ago
I downloaded some 'uncensored' local models around the beginning of this year.
Their furry porn is crap, or maybe I'm just not into that. But they generate it at least.
However, the answers to technical questions are a lot more concise and to the point, which is far less annoying than the big names.
Haven't bothered updating the models though, so now I drifted back to Gemini for quickie API questions.
Funnily enough, of all that I've tried, the model by the best at writing porn has been not one of ones uncensored and tuned exactly for that purpose, but stock Command R - whose landing page lists such exciting uses as "suggest example press releases" and "assign a category to a document".
> uncensored and tuned exactly for that purpose
Are they tuning too, or just removing all restrictions they can get at?
Because my worry isn't that I can't generate porn, but that censorship will mess up all the answers. This study seems to say the latter.
Usually "uncensored" models have been made by instruction tuning a model from scratch (i.e. starting from a pretrained-only model) on a dataset which doesn't contain refusals, so it's hard to compare directly to a "censored" model - it's a whole different thing, not an "uncensored" version of one.
More recently a technique called "orthogonal activation steering" aka "abliteration" has emerged which claims to edit refusals out of a model without affecting it otherwise. But I don't know how well that works, it's only been around for a few weeks.
2 replies →