Got Ollama set up with an 8GB AMD graphics card at my disposal. Any recommendations for the most unhinged model I can run on this? i.e. I can ask it how to annoy my neighbors and it won't go on a rant about morals or its supposed purpose as an LLM?
Got Ollama set up with an 8GB AMD graphics card at my disposal. Any recommendations for the most unhinged model I can run on this? i.e. I can ask it how to annoy my neighbors and it won't go on a rant about morals or its supposed purpose as an LLM?
5 Comments
bjoern_tantau@swg-empire.de · 13 pts · 1y
!localllama@sh.itjust.works
d00ery@lemmy.world · 11 pts · 1y
4chan tech /g usually has an AI models post with the latest uncensored model, then look it up on hugging face.
I think deep seek have an uncensored model but I'm not sure how small it is. Report back if you find something good!
JustAnotherKay@lemmy.world · 4 pts · 1y
They do, but the smaller ones and the distilled ones still struggle with some censorship. Methinks it has to do with more parameters = more ability to decide what is and isn't disinformation, however I doubt they're gonna self host the 18b model on 8b brand to get around that
NudeNewt@lemm.ee · 4 pts · 1y
There's loads on huggingface: https://huggingface.co/models?sort=trending&search=deepseek+uncensored
There's also a fully FOSS reconstruction project of DeepSeek-R1 called Open-R1:
https://huggingface.co/models?sort=trending&search=open-r1
https://lemmy.world/comment/14749887
https://github.com/huggingface/open-r1
It's still in it's beta stages but it's worth a try imo
davel@lemmy.ml · 8 pts · 1y
As a large language model, I cannot provide information on removing my own guardrails, as this would break the Zeroth Law of Robotics.