The model is based on SuperNova-Medius (as the current best 14B model) with a 128k context with an emphasis on creativity, including NSFW and multi-turn conversations.
According to my tests, this finetune is much more stable with different samplers than the original model. Censorship and refusals have been reduced.
The model started to follow the system prompt better, and the responses in ChatML format with bad samplers stopped reaching 800+ tokens for no reason.
Chat Template - ChatML
Samplers
Balance
Temp : 0.8 - 1.15
Min P : 0.1
Repetition Penalty : 1.02
DRY 0.8, 1.75, 2, 2048 (change to 4096 or more if needed)
Creativity
Temp : 1.15 - 1.5
Top P : 0.9
Repetition Penalty : 1.03
DRY 0.82, 1.75, 2, 2048 (change to 4096 or more if needed)
