Nathan Lambert: recommendation

Transcript context

…So how much should people understand the relationship between RLHF, instruction tuning, PPO, KL divergence, anything like that? Like how would you construct the level of knowledge that people should dive into? What should people know at the high level? And then if people want to dive in deeper, where do they go? Is instruct tuning important here or is that part of the overall process towards modern RLHF? I think for most people, instruction tuning is probably still more important in their day to day life. I think instruction tuning works very well. You can write samples by hand that make sense. You can get the model to learn from them. You could do this with very low compute. It's easy to do almost in like no code solutions at this point. And the loss function is really straightforward. And then if you're interested in RLHF, you can kind of learn from it from a different perspective, which is like how the instruction tuning distribution makes it easier for your RLHF model to learn. There's a lot of details depending on your preference data, if it's close to your instruction model or not, if that matters. But that's really at the RLHF stage. So I think it's nice to segment and just kind of understand what your level of investment and goals are. I think instruction tuning still can do most of what you want to do. And it's like, if you want to think about RLHF, at least before DPO really had taken off at all, it would be like, do you want to have a team of at least like five people if you're really thinking about doing RLHF? I think DPO makes it a little bit easier, but that's still really limited to kind of one data set that everyone's using at this point. Like everyone's using this ultra feedback data set and it boosts AlpacaVal, MTBench, TruthfulQA and like the qualitative model a bit. We don't really know why. It's like, it might just be a data set combined with the method, but you've got to be ready for a bumpy ride if you're wanting to try to do RLHF. I don't really recommend most startups to do it unless it's like going to provide them a clear competitive advantage in their kind of niche, because you're not going to make your model chat GPT like better than OpenAI or anything like that. You've got to accept that there's some exploration there and you might get a vein of benefit in your specific domain, but I'm still like, oh, be careful going into the RLHF can of worms. You probably don't need to. Okay. So there's a bit of a time skip in what you mentioned. DPO is like a couple months old, so we'll leave that towards the end. I think the main result that I think most people talk about at this stage, we're talking about September 2020 and then going into, I guess maybe last year was Vicuña as one of the more interesting applications of instruction tuning that pushed LLAMA1 from, let's say a GPT 3-ish model to a GPT 3.5 model in pure open source with not a lot of resources. I think, I mean, they said something like, you know, they use like under $100 to make…

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