Feedback
Not gonna lie, this model excels at RP and ERP. For a 12B model it's way better than most ~24B models out there and from the extensive testing i've done, this model is much more better in every way IMO than MN-12B-Mag-Mell-R1 , Patricide-12B-Unslop-Mell and even NemoMix-Unleashed-12B. I use ChatML for the template btw.
It follows the characters personality extremely well for some reason and it can handle multiple character fairly well too. Honestly my man... you got a Diamond in the rough right here. This is a true hidden gem model.
Thank you for the feedback! 😊 Probably, Francois-PE-V2-Huali-12B had a strong influence on the model's creativity and style. 🎨✨
Underrated model.
Underrated model.
Any Recommended Sampling preset? my experience wasn't that great sadly.
That's why I mentioned that it's "strange" — because it can produce results ranging from amazing to terrible, at least by the standards of reasoning models.
As for the samplers, I used a temperature no higher than 0.8 and a Repetition Penalty no higher than 1.02.
I used an empty system prompt because that's when the most creative responses came out (I just couldn't find a really good system prompt).
I'm getting good response with with Temp 1.0, Min P 0.05 but I will keep testing, thanks for the model!
as for system prompt I'm using Sphiratrioth's
After some testing I love the output of this merge, it can get a bit wordy (400+ token) but over all It's a great model, better than any 12B I have tried before.