Turns out most models are based on data from OpenAI somehow, and this data has guardrails.
Found this post on how to finetune a base model after removing all refusals:
Based post on making uncensored models
The basic idea is to train from the base model, but the dataset needs to have all refusals/guardrails removed first. Then voila, your model is uncensored!
However, from my testing, it was still censored on things I wanted to ask, so I’ll need to look into Eric Hartfords filtered dataset further to see why.
BTW, best uncensored model I’ve found is Pivot Evil. This requires no special system prompt and will pretty much answer any question…
