just an "abiliterated model"
#4
by damejugg - opened
I'm not sure how you got the KL Divergence value you gave in the model card because but this model is diverging so much the way when attaching adapters/loras to a model. it's great in regard to minimizing refusals, but in turn it seems it achieve that by framing the answers as a "fictional story" or something, and it does that to every request whether it triggers a refusal or it is a normal request/question.
I actually like it but I don't think I could call it an "abiliterated gemma-4 model" , it's more like just an "abiliterated model"