If you do a version 1.3 ...
If there is a vesion 1.3, please implement some safeguards to prevent it from speaking on behalf of the user. It ignores the model instructions and anything you write, continuing to speak on the user's behalf regardless of what you do.
I restarted and deleted the chat several times, but unfortunately it became too annoying to continue, so after about 30 minutes, i deleted it, even if the overall writing was kinda nice.
I don't have a method for testing this feature (user impersonation). It could be a result of the karcher merge method or one of the components
There is a v1.3 and soon v1.4 but I don't know if these would fix that. If you had a chat json you could upload from kobold where it recreates this user impersonation I could run it through some of the component models to see what might be causing this
Basically I tested for censorship and creativity, and with Mistral tekken i usually dont have this issue BUT when its trying to control multiple chars i saw it happen more it could be something with mistral in general
Mistral itself generally has no issue with speaking for the user. I mostly use Mistral due to the lack of proper grammar in most of the other models (e.g. 'open door, walk out' instead of 'I open the door and walk out'), which drives me crazy. In most cases, if the model is instructed with something like 'You are strictly forbidden to write, think or act for {{user}}!', it may go too far in isolated instances. However, if you correct the output and delete the impersonation, it usually works. The issue is if the model was trained with text that is not only talking from {{char}} POV, but also from the {{user}}'s POV, then the model sees this behaviour as normal and will keep doing it, no matter how often you tell it not to. But with Qlora or Lora training there are ways to even with only a hand full of examples reduce such behavior ... even chatgpt can help there ;) ... this is at least my limited knowledge on it ;)