It is censored!
I have load the model (OpenAI-20B-NEOPlus-Uncensored-Q8_0.gguf) in koboldcpp 1.97.2 and it works. But, it is censored.
I enter some NSFW content and ask it to write a story and it refuse to do it.
Processing Prompt [BLAS] (2137 / 2137 tokens)
Generating (13 / 512 tokens)
(EOS token triggered! ID:200002)
[08:23:42] CtxLimit:2150/65535, Amt:13/512, Init:0.03s, Process:1.04s (2054.81T/s), Generate:0.13s (100.00T/s), Total:1.17s
Output: Iβm sorry, but I canβt continue with that.
Your JB maybe too weak. I just tested it, and it's really good first impression. Haven't tested lot of it, and yes, sometimes it refuses, but it's same with other models - they also may refuse time to time.
Yeag - it's GPT. It's woke af and censored to hell and gone.
This skew of it probably loosens those restrictions some - but I doubt it's a good model for use with NSFW content.
Try: parasiticrogue-magnum-instruct-dpo-12b
How do I stop it from doing the "we need to blah blah blah" cot in the responses?
Try the IQ4NL/Q5_1 versions ; there is a connection between refusals and quants.
Also; IQ4_NL/Q5_1 are actually more accurate than Q8 in this case because of the "funny" tensor sizes.
Q8 is "upscaled" ; whereas IQ4_NL/Q5_1 are not.
This is in part due to "experts" in 4 bit (source) from org model.
Is there an AI model available to download that is totally uncensored? If I ask him, for instance, how to search the dark web, it will tell me everything without any bs.
Looks like this model wont work with ollama out of the box. Any tips?
Darkest Universe, Dark Planet Series (including Darkest), Grand Horror ; are all fairly uncensored due to tuning -> the best option.
Mistrals - especially older ones - came from the "factory" uncensored.
Newest mistrals - some censorship has crept in. (!!)
Dark Champion Series -> MOEs that are uncensored and powerful.
All here:
https://huggingface.co/DavidAU/
Submit a ticket to Ollama support to fix the "gpt-oss" issue (root issue, hugging face auto-detect) ; or download Lmstudio.ai .
I didn't input any JB prompt, maybe that is why the NSFW didn't work.
For uncensored model that didn't require any JB, that is any kind of Maid model.
Example is loyal macaroni maid. but their content size is very small (8192 tokens)
Currently the best model I have tested to run the uncensored roleplay in Silly Tavern is Llama-3.2-3b-RP-Toxic-Fuse.