I'm curious what the reason for including this, because at the end of the day its a string of text (tokens) going into a function and it spitting out another string of text.
Maybe they are using the prompts and responses for training doesn't want these to influence future model behavior?
This is my assumption. I can't think of any other sensible explanation.
My opinion: https://news.ycombinator.com/item?id=50012423