Greatly surprised by this model.
Hear me out. This model can be prompted into anything and it CAN write quite long and coherent story paragraphs in RP chat.
One of the most positively shocking findings was a good understanding of 'empathy' by this model (test case: RP scenario where user speaks of some cruel stuff presenting it as "hey {{char}}, have you ever thought of ????", then receives character's output, and in a follow-up message the user shows a revelation about his traumatic past, indicating that the aforementioned horror has actually happened to him and it still haunts him (PTSD) - ultimately, most LLMs fail to handle such case, making the character either feel offended in both responses due to the sheer explicitness of suggested horrors, or making the character probe into user invasively (demanding more details on horrific trauma), or even suggesting to mirror/re-enact the said cruelty... BUT THIS ONE HANDLED IT PERFECTLY, making the character offer warmth and comfort to user - and that wasn't just a lucky spark of empathy, it handled the case consistently over multiple regenerations).
Anyway. This IS a gem. Mad respect, Mr. Drummer. You've done something remarkable. It's probably not goint to work so good without an appropriate prompt, but nonetheless it IS brilliant. Thank you.
Are you using this with SillyTavern? How were you able to get it to work? I'm having trouble with it.
Are you using this with SillyTavern? How were you able to get it to work? I'm having trouble with it.
Yes, KoboldCPP + ST. Pardon me if whatever I wrote was overly optimistic - it's tricky to work with, residual Gemma3 issues are lurking beneath (like poisoning the narrative with metallic smells and 'disgust' if the model senses something 'unethical'). Frankly, I feel like I've already messed it up by tinkering with settings I didn't bother to save ._.
Pretty sure I've had it set to something like: Temp=1, Top_p=0.95, Top_k=64, Min_p=0.01 and Rep_pen=1.05 (range 1024) while experimenting with tfs (0.85 - 1) because the latter seemed to affect how the model falls into endless generation. Sampler order 6,0,1,3,4,2,5 originally, then 3, 6, 0, 1, 2, 4, 5 --- vastly different results (first one tended to write long and verbose paragraphs, second one was more like 'normal' Gemma with all its excessive moralization - probably no longer able to show 'empathy' I spoke of).
So, for RP all stars need to align, meaning a prompt that LOGICALLY convinces the model (something along the lines of 'we're now on another planet, {{char}} follows this and that' -- without pressuring it with prohibitions). If the character's profile contradicts with user's input, the model will likely default back to the aforementioned Gemma3 issues.
Bottom line, it can either be outstanding or a pain in the ass. It's certainly not a universally applicable NSFW RP model. Looking back, it's as if the model 'glitches out' sometimes in a strange, beneficial way - only when correct conditions are met, like the character having a personality that does not mind (according to AI's logic/values/morals) whatever wild thing the user had suggested.
like poisoning the narrative with metallic smells and 'disgust' if the model senses something 'unethical'
@AutisticPancake Try https://huggingface.co/BeaverAI/Gemma-3-R1-27B-v1a-GGUF It's much more positive but it won't spiral into negativity I think. You'll probably need to do a lot of prompt wrangling to make it evil.