Gemma
Are there any plans to work on gemma settings?
Nothing firm, but I do intend to look at the model once issues with context beyond 4k tokens during inference are dealt with.
I've uploaded some settings to try for Gemma2. It should significantly reduce refusals, although warnings and disclaimers can still pop up. Gemma2 censorship seems really deeply embedded. Tested against SPPO Iter3.
TYVM - playing with it now, and I will let you know if I come across any issues.
So far, the only issue I am having is that it has difficultly impersonating user. I'm using it paired with your Smoothie sampler.
I've uploaded a second version, reducing the token cost of the Instruct template.
I used it last night, impersonations seem to be much better, and didn't have any issues with it. The over all play pretty unique, and quite a bit of fun vs. L3.
Give v3 a try. It's smaller, but immersion should be deeper.