[robviren]

My early attempts at uncensoring the models sort of stumbled into this as well. I told the models to produce output heavily using the transposed letter effect. I would also have it remove random letters in words longer than 4 letters. The end result was a broken, readable, but surprisingly consistent output.

I wrote a little script that would apply a regex to scramble words and remove letters to ask questions. The models could almost always use this input which was crazy to me. Made for some fun time having uncensored conversations in complete nonsense with the models.