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

[–] venusaur@lemmy.world 0 points 2 weeks ago (1 children)

Thanks! I don’t think I can run an 8B yet. Need to invest in a better machine. I’m stuck on 4B Q4.

The uncensored Qwen that I’m using started throwing infinite ?’s at me one time. Had to restart it and has been fine since.

[–] Rhaedas@fedia.io 1 points 1 week ago (1 children)

It is certainly inaccurate, but in my mind's picture of how the transformers work, reducing their quantization and also doing what abliteration does, there is a line where you've done a lot of "damage" to the original model and so there will be places where it just hangs or goes off on severe tangents. There are good places for even the 1bit models where they don't get pushed to hard, but there are limits for them all, including the big ones.

Hugging Face does have a few Q4_K_M versions. Maybe something will fit.

[–] venusaur@lemmy.world 1 points 1 week ago

For sure. You would need a model that is not censored at training.