Sup! Some thoughts.
#2
by
MateoTeo
- opened
I played around with the default Qwen2.5 32B, and must say that this finetune tries to avoid many themes too.
Just like RP with the original Instruct model, it tries to evade or moralize, changing characters' personalities, and ignoring pretty good system prompts and examples. Still feels pretty dry as the original and the themes are not even so extreme.
ParasiticRogue/EVA-Instruct-32B managed to deal with it by 40/60 merge of Instruct and RP-trained finetune (non-instruct, uncensored) - that model flows nicely. A bit dumber, maybe a 50(or 60)/50(40) would be a more sweet spot. Perhaps that will help, good luck!