Eval Numbers
Did you checked eval numbers before/after uncensoring?
Did you checked eval numbers before/after uncensoring?
I always do some quick llama.cpp evals over ARC Easy, ARC Challenge, MMLU and Winogrande before and after finetuning to make sure I don't break the model. The eval results seamed relatively unaffected so far. I will do more careful measurements which I will share once I'm done reasoning finetuning this model because in my opinion testing a reasoning model on multiple choice questions without giving it the opportunity to think is bad practice and in its current state this model obviously lost most of its reasoning capabilities. Unfortunately reasoning finetuning is at least 10 times as resource intensive as Uncensoring or Unbiasing this model so it will still take a few days for it to be done training.