Alessio Fanelli: commitment

Transcript context

…Models all the way down. think also the other, I mean, to me, some of these things are like chemy, right? They're like, we stopped annotating super fast and fine tuning data at 27,540 annotations. Why? It's like, it seems like such a arbitrary number, you know, that I feel like that's gonna be one of the next research areas, you know, figuring out where the, the right limit is. Do we have maybe, do you know if there's any really good again, like open source? Open source, like datasets for posts, not pre-training, but like a fine tuning then R lhf. Because I think one of the big moments with Uber pajama was like, okay, we can take the LAMA one data mixture, use all the open source data sets and just run GPUs at them. How do we get to do the same with the post-training flow? Okay, so you were breaking up a little bit for the question. So I, I'm gonna say what I think it was, and if it wasn't, you can jump in and clarify. So I think it's like, how do we recreate this supervised training data set and like, can we do anything else with it after the fact? Yeah. So Gen, this is another thing that we've started doing, and I think that what, so the open source equivalents are something like Open Assistant created a really high quality dataset, artifact, and then the recent trend is for this thing that's like called Uncensored dataset, which I think is this totally silly name. Because really what they're doing is they're removing instructions like as a language model, I don't wanna say this. And therefore when you remove these things, the model gets more helpful. So that's just gonna be the new type of data, which is just clean response on instructions with really strong distribution control. And the thing is about recreating this is that it's. Hard to create a diverse set of tasks. So what they are essentially paying money for is someone to make sure that you're not getting a whole bunch of the same poems or something. It's like getting 27,000 weird creative tasks that don't all overlap with each other is why you have to pay a lot of money for it. Rather than saying, oh, we have 250 people on this call, it's all due, 10 of them. And then that's a solid start. Like we would just have a totally misshape in distribution and it wouldn't be that useful. So I think even in, so you can go look at like instruction, BT and other papers like this have breakdowns of what that instruction data, the supervised, fine tuning data actually looks like. But actually creating it is pretty hard. And I do think that the vendors provide a really high quality amount of data, but their point about the models being able to create it is also really true. So it's, it's, it's pretty borderline right now. And anthropic stop using that in their, in their future work. So like, Philanthropics new base models are just good enough at responding to instructions where they don't need to do supervised, fine tuning. And that's like in the constitutional AI paper. So it's like, I don't think that's the place to invest time. It's much more on the preference side to get the RL HF model and to get these preference models going. So then maybe you can even do creative things like constitutional AI and stuff after that.…

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