A few words..
"...where characters are portrayed creatively and proactively." what do you mean by that? Any sample cards?
I've been testing the model for two hours with the recommended settings, and it's very solid. I had the promt written in XML format β your model doesn't seem to like it (I'm not sure about XML character sheets β any suggestions?). The longest roleplay is over 16k and still stable.
Top nsigma 0.9 also performs interestingly (more creative and stable), but I don't think DRY Repetition Penalty works then, there are some repetitions - not many.
Hey!
Below is a couple basic example of what I consider "creatively & proactively". Basically, the bot should drive the conversation forward while responding and give you plot hooks to work with, rather than require that all the decisions and ideas come from your side.
Most cards should work. There's cards written in markdown format, plain text, ali:chat and XML within the datasets used. Although the predominant format is markdown. It wouldn't surprise me there's some cards it struggles with though (still a comparatively small local model at the end of the day). If using XML my suggestion would be try to minimize the tags, for example using <char></char>
with plain text inside for the appearance, personality etc.
The samplers certainly aren't fixed or a hard requirement. They're more just there for those who don't want to experiment with them or as a baseline to build your own settings with.