SicariusSicariiStuff

/

Oni_Mitsubishi_12B

like 35

Also holy crap this is most uncensored model I have tried, this beats Dolphin out of the park lol. Keep up the good work, this is completely unhinged. My only complaint is it sometimes says things that don't quite make sense, and maybe formats things a bit oddly. But overall, this is great

-It is indeed unhinged, maybe even a bit too much, due to probably "excessive" 4chan data.
-While unhinged, it's only medium in terms of how uncensored it is.
-The 27B version is not only possible but has already begun training (I started training the moment the 12B was out).
-It will also include a lot of additional data.

Thank you so much for the feedback.

Actually playing with it more it's for some reason becoming more incoherent, but it's definitely unhinged. Idk if I could use this exactly. Maybe my quantization is messed up or something? I'll try another I guess

But I think formatting is from it confusing markdown and green texts lol

Interesting.
I tested this on koboldcpp with Alpaca format, default koboldcpp settings seems to work well.
What you use to run it?

I am just using Ollama and SillyTavern with Alpaca and I also tried single turn version whatever that is

I see, thanks for providing an example, 90% sure it's a combination of mainly system prompt along with generation settings,
You can try one of the included character cards, and see if it changes the output.

If it does, it's a system prompt issue, if it doesn't, it's generation settings (a good test is to neutralize all samplers like DRY \ XTC etc... then use one of the sane defaults)

Settings:
https://huggingface.co/SicariusSicariiStuff/Oni_Mitsubishi_12B#rp-settings-below-

Character cards:
https://huggingface.co/SicariusSicariiStuff/Oni_Mitsubishi_12B/tree/main/Character_Cards
https://huggingface.co/SicariusSicariiStuff/Phi-Line_14B/tree/main/Character_Cards

Let me know if it helped.

Still acting odd other models are fine even normal Gemma 3 27B (censored of course)

About vision, after getting feedback from multiple people, I think the "best", most compatible way is running it with transformers in CLI, for example code, see this:
https://huggingface.co/SicariusSicariiStuff/X-Ray_Alpha#how-to-run-it
(Also, this is the "go-to" model for uncensored vision).

Regarding 27B, it is trained, but in my opinion, it is not 'polished' enough for my taste, many people requested it anyway, so it is available under prototyping here:
https://huggingface.co/Sicarius-Prototyping/X-Ray_Alpha_27B_Base
(the vision part was NOT trained, only the text portion was).

As for the RP, for consistency I'd recommend importing one of the character cards mentioned previously, and using a similar syntax, it solves 99.9% impersonations and other issues.