Alistair Pullen: evaluation

Transcript context

…I just wanted to dive in a little bit more on your experience with the fine tuning team. John Allard was publicly sort of very commentary supportive and, you know, was, was part of it. Like, what's it like working with them? I also picked up that you initially started to fine tune what was publicly available, the 16 to 32 K range. You got access to do more than that. Yeah. You've also trained on billions of tokens instead of the usual millions range. Just, like, take us through that fine tuning journey and any advice that you might have. It's been so cool, and this will be public by the time this goes out, like, OpenAI themselves have said we are pushing the boundaries of what is possible with fine tuning. Like, we are right on the edge, and like, we are working, genuinely working with them in figuring out how stuff works, what works, what doesn't work, because no one's doing No one else is doing what we're doing. They have found what we've been working on super interesting, which is why they've allowed us to do so much, like, interesting stuff. Working with John, I mean, I had a really good conversation with John yesterday. We had a little brainstorm after the video we shot. And one of the things you mentioned, the billions of tokens, one of the things we've noticed, and it's actually a very interesting problem for them as well, when you're How big your peft adapter, your lore adapter is going to be in some way and like figuring that out is actually a really interesting problem because if you make it too big and because they support data sets that are so small, you can put like 20 examples through it or something like that, like if you had a really sparse, large adapter, you're not going to get any signal in that at all. So they have to dynamically size these things and there is an upper bound and actually we use. Models that are larger than what's publicly available. It's not publicly available yet, but when this goes out, it will be. But we have larger law adapters available to us, just because the amount of data that we're pumping through it. And at that point, you start seeing really Interesting other things like you have to change your learning rate schedule and do all these different things that you don't have to do when you're on the smaller end of things. So working with that team is such a privilege because obviously they're like at the top of their field in, you know, in the fine tuning space. So we're, as we learn stuff, they're learning stuff. And one of the things that I think really catalyzed this relationship is when we first started working on Genie, like I delivered them a presentation, which will eventually become the blog post that you'll love to read soon. The information I gave them there I think is what showed them like, oh wow, okay, these guys are really like pushing the boundaries of what we can do here. And truthfully, our data set, we view our data set right now as very small. It's like the minimum that we're able to afford, literally afford right now to be able to produce a product like this. And it's only going to get bigger. So yesterday while I was in their offices, I was basically, so we were planning, we were like, okay, how, this is where we're going in the next six to 12 months. Like we're, Putting our foot on the gas here, because this clearly works. Like I've demonstrated this is a good, you know, the best approach so far. And I want to see where it can go. I want to see what the scaling laws like for the data. s here, because this clearly works. Like I've demonstrated this is a good, you know, the best approach so far. And I want to see where it can go. I want to see what the scaling laws like for the data. And at the moment, like, it's hard to figure that out because you don't know when you're running into like saturating a PEFT adapter, as opposed to actually like, is this the model's limit? Like, where is that? So finding all that stuff out is the work we're actively doing with them. And yeah, it's, it's going to get more and more collaborative over the next few weeks as we, as we explore like larger adapters, pre training extension, different things like that.…

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