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

[–] e0qdk@reddthat.com 2 points 5 days ago

Thanks for the tips. That sounds similar in performance to what I'm seeing, so I probably didn't screw up too much trying to get it working. If you're using it in more of a story writing capacity than a chat capacity, that makes sense.

You may want to set a system prompt

I tried initially with ollama run on the command line just to see if it was working at all when I got that response. (It amused me, so it stuck with me.) I've tried again with my custom tooling -- which does set a system prompt (geared more towards assistant style ussage though) -- and it didn't really take anything from the prompt. It's possible I don't have something set up right with templates, but I'm probably going to shift over to llama-server eventually anyway...

Following your suggestions on system prompt style though I was able to get it to give me a more specifically targetted coherent story via llama-cli. If I poke at it a bit, I'll probably figure out some use for it. It's pretty creative.

If you're curious about my findings from the uncensored qwen3.6 I mentioned, it generates pretty quickly on my machine (~50 tok/s give or take 5 depending on quantization) and I haven't gotten it to outright refuse anything yet -- but I've only poked at it a little. Based on the other comment in the thread here about llama penises, I whimsically asked it to "Generate a sexually explicit song about llama penises." and it did without complaint. (Stock qwen3.6 refused, of course.)