These models were finetuned on synthetic chats (as opposed to documents) to have edited beliefs. Note: no pretrained data mixed in.