This is about RLHF training. But I've wondered if something similar could be used to automatically judge the quality of the data that is used in pre-training, and then spend more compute on the good stuff. Or throw out really bad stuff even before building the tokenizer, to avoid those "glitch token" problems. Etc.