Based on Mistral 7B. Instruction format is ChatML. Good at NSFW. Good at following character cards and instructions, considering parameter count.
Very good RP model, especially for 7B. I tried Q5_K_M and Q8 GGUF quantizations, didn't notice much difference. It's a bit autistic for being a finetune of the praised Mistral, but it's soulful and fun. NSFW descriptions are great, a lot of dialogues in the replies. Not very verbose, but adequate with the descriptions of scenes and actions. Though keep in mind that the model is too dumb to be used for adventure/story based RP. In my experience, the model works best when I cut off very poor tokens with topP 0.95, then apply the temperature a bit higher than 1, and after that throw away another tail of relatively bad tokens with TFS of 0.9-0.95.