I downloaded an uncensored aggressive Qwen 3.5 model and I can see in its reasoning that it is still limiting responses based on safety guardrails (e.g. violence, NSFW).

Anybody have recommendations for truly uncensored models?

EDIT: I turned off reasoning and I think it’s more uncensored if I’m very specific about what the response should include.

you are viewing a single comment's thread
view the rest of the comments

[–] hendrik@palaver.p3x.de 3 points 6 months ago* (1 child)

I didn't have any luck with some uncensored Qwen 3.5 either. It always reasons about the guardrails. And it leans towards weaseling itself out of the situation. And the 3.5 version goes on for 1500 tokens anyway, just to think about how to respond to "Hello".

I didn't do a lot of LLM stuff lately. I'm also looking for a new local model which isn't censored nor a sycophant, nor overly verbose and repetetive. But I guess I see that with a lot of models. And lots of the supposedly uncensored ones will give you the kids version of a murder mystery story, because they're still averse to violence, conflict, taboo and all kinds of things.

And a lot of internet recommendations are older models from at least a year ago?! At least I didn't find any perfect fit (yet).

permalink

fedilink

source

hideshow 1 child comment

[–] venusaur@lemmy.world [S] 2 points 6 months ago

The reasoning for hello is crazy haha. I’ve experienced the same, but if you turn off reasoning on launch and explicitly state the rules you want it to break I’ve had some success. I was trying to get it to tell me a story about llamas having sex and it went on forevvver reasoning about why it shouldn’t say things and how to rephrase to not break rules. The funniest part of the reasoning was “llamas don’t have penises (obviously, they’re mammals)”. Haha it reasoned itself into thinking llamas, and mammals, don’t have penises.

permalink

fedilink

source

parent