My guess is you'd better use your message + a bunch of other messages with RLHF, yours being the positive one and the others negative. If you don't have other m

My guess is you'd better use your message + a bunch of other messages with RLHF, yours being the positive one and the others negative. If you don't have other messages, just generate some. You can also do some supervised fine tuning to have the model memorize your facts.