Suggested sampling settings
Hello. First of all, thank you for tuning this model.
What sampling settings (temperature, TopP, etc) do you recommend for this model and for 30B one?
Offtopic question:
Are there any other RP sites, that can be used for finetuning further such models?
I don't have any particular settings I highly endorse, since the model is still WIP and the settings may vary according to preference, but I might suggest starting with temperature something like 0.9, and increase it from there a bit in increments of 0.05 or so depending whether you find it too robotic or dull. For top_k and top_p, try 40 and 0.5, respectively. The repeat penalty tokens should be around 100-300, with the penalty around 1.1~1.2.
Are there any other RP sites, that can be used for finetuning further such models?
There certainly are, and there are some scrapes available from those from which datasets can be further derived:
https://rentry.org/qib8f
Quite many might make an interesting addition, and perhaps help further pushing the context size as well.
If you're interested in including anything from any in to an RP finetune cocktail, contributions are welcome by cleaning the samples for instance. There's a handy tool from @Squish42 now that can make the job easier, and one can check out here https://huggingface.co/datasets/Squish42/bluemoon-fandom-1-1-rp-cleaned/discussions/1#646ed5b034fde71fdaa69e5c as an example how the bluemoon fandom was prepared.