this third time i try rewrite this introduction on this thread, while i am not sure why i need do this i feel is important to make good introduction. So hello everyone good night from here i hope you enjoy your holiday or thanksgiving.
Uncensored Local Ai model is love by everyone and championed like golden kid in some case (mistral for example) because it just good compare it align version of it. So there thinking that uncensored model are fully uncensored is completely bias and not true.
Some model are creation of various model that being merged to create new model that not align but retain it capability and value, such as Mistral-Orca 7b, other are just being lora or fine tune with dataset that uncensored but still the based model are still not uncensored but align one, make it stil generate output advising or lecturing user instead give straightforward answer.
model can be called truly uncensored when it not given user advice and act faithfully generate user desire output.
thats my thought for this time, i hope everyone enjoy your thanskgiving.
You mean, the later fine-tuning is tainted by the censored base model?