Alessio Fanelli: belief

Transcript context

…es to generate, punctuation, weird tokens like calculator tokens. It could overfit to anything if it's in the data a lot and it happens to be in a specific format. And the KL constraint prevents that. There's not that much documented work on that, but there's a lot of people that know if you take that away, it just doesn't work at all. I think it's something that people don't focus on too much. But the objective, as I said, it's just kind of, you optimize the reward. The reward is where the human part of this comes in. We'll talk about that next. And then subject to a constraint, don't change the model too much. The real questions are, how do you implement the reward? And then how do you make the reward go up in a meaningful way? So like a preference model, the task is kind of to design a human reward. I think the equation that most of the stuff is based on right now is something called a Bradley-Terry model, which is like a pairwise preference model where you compare two completions and you say which one you like better. I'll show an interface that Anthropic uses here. And the Bradley-Terry model is really a fancy probability between two selections. And what's happening in the math is that you're looking at the probability that the chosen completion, the one you like better, is actually the better completion over the rejected completion. And what these preference models do is they assume this probability is correlated to reward. So if you just sample from this probability, it'll give you a scalar. And then you use that reward later on to signify what piece of text is better. I'm kind of inclined to breeze through the math stuff because otherwise, it's going to be not as good to listen to. I think people want to hear it. I think there's a lot of higher level explanations out there. Yeah. So the real thing is you need to assign a scalar reward of how good a response is. And that's not necessarily that easy to understand. Because if we take back to one of the first works, I mentioned this tamer thing for decision making. People tried that with language models, which is if you have a prompt in a completion and you just have someone rate it from 0 to 10, could you then train a reward model on all of these completions in 0 to 10 ratings and see if you can get chat2BT with that? And the answer is really kind of no. Like a lot of people tried that. It didn't really work. And then that's why they tried this pairwise preference thing. And it happened to work. And this Bradley Terry model comes from the 50s. It's from these fields that I was mentioning earlier. And it's wild how much this happens. I mean, this screenshot I have in the slides is from the DPO paper. I think it might be the appendix. But it's still really around in the literature of what people are doing for RLHF.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.