this is a test.... trying to get Qwen3 30B to behave.
I started by fine-tuning SuperbEmphasis/Black-Eclipse-Test-ERP-RP-v2 with my improved dataset.
however this time, I cranked the activated experts to 64/128 to force more parameters to get trained.
Then I cranked the experts back down to 24/128.
it is surprisingly coherent. the biggest issue I have run into so far is that it REALLY doesnt like roleplay scenarios where '{{char}}' can be multiple characters. It wants to stay as the single initial characters This tells me that I need more scenario based examples...