Alistair Pullen: evaluation

Transcript context

…history of this is, is, is quite simple, actually. SweeBench, up until, I want to say two weeks ago, but it might be less than that, or more than that. But I think two weeks ago, suddenly started mandating what they call trajectories, when you submit. So, but prior to this, essentially, when you run SweeBench, you run it through their harness, and out the other end you get a report. json, which is like, here's how many I resolved, here's how many I didn't resolve, these are the IDs, the ones I did, these ones the IDs I didn't, and it gives you any ones that might, might have errored, or something like that. And what you would submit would be all of your model patches that you outputted and that report. And then you would like PR that into the sweep entry per and that would be it. That was the still the case when we made our submission on whatever day it was. They look at them every Monday. We submitted it at some point during the week. I want to say it was for four days before that. And, um, I sort of like sat back and waited. I assumed it would be fine when it came to Monday. Um, they then said, actually, no, we want model trajectories. And I was like, okay, let me see what this is. And so on. I sort of dug into it and like model the trajectories are essentially the context window or like the reasoning process of like, show you're working. How did you get here? If you do a math exam, show me you're working. Whereas before they were like, just give me the final answer. Now they want to see the working, which I completely understand why they want to see that. Like the SWE bench fundamentally is an academic research project and they want all the stuff to be open source and public so people can learn from each other and improve and so on and on. Very good. I completely agree. However, at least for us, and the reason that we are not on the leaderboard is that obviously the model outputs that we generate are sort of a mirror of our training data set, right? Like you train the model to do a certain thing and output a certain way. Whatever your output looks like, your training data for the moment, as a closed source company, like fighting for an Edge, we've decided not to publish that information for that exact reason. I don't want someone basically taking my tra. And then taking a model that's soon going to be GA and just distilling it immediately and then having genie for themselves. And, you know, as a business owner, that's the decision I've had to make. The patches are still public. So like the, dare I say, traditional SweeBench submission, you can go to our GitHub repo and see it and run them for yourself and verify that the numbers come out correctly. Like that is all, that is the potted reason. That's the story. That's the story. Uh, SweeBench verified. You have a score. I do have a score. I do have a score. 43. 8%? the numbers come out correctly. Like that is all, that is the potted reason. That's the story. That's the story. Uh, SweeBench verified. You have a score. I do have a score. I do have a score. 43. 8%? It's one of those things where like there aren't that many people on the leaderboard yet, so you don't know how good or bad that is. And it's smaller data set, right? Oh, it's, it's great. So on a tangent, Swebench, original Swebench was 2, 294. Which is expensive. It's like 8, 000 to run. Oh, that's cheap. That's cheap, what are you talking about? I don't know, at least for us, I don't even want to say publicly how much it cost us. How much it cost us to run that thing. Expensive, slow, really like crap for iteration, because like, you know, you make a change to your model, how does it do on SweetBench? I guess that's why SweetBench Lite existed, but SweetBench Lite was not a It was, it was easy stuff, right? It wasn't a comprehensive measure of the overall thing. So we actually had the idea a month ago to, what we were going to call SweeBench Small, where we were going to try to map out across SweeBench, like, what is the distribution of, like, problem difficulty and all these different things, and try to come up with, like, 300 examples that sort of map that, where, you know, Given a score on SWE Bench more, you could then predict your SWE Bench large score and sort of go from there. Fortunately, OpenAI did that for us, and probably much better than we would have done. They used some human labelers, and as obviously we're working with OpenAI quite closely, they talked to us about it, and they, Um, you know, we're able to let us know what the instance ID were, IDs were that were in the, the new suite bench version. And then as soon as I had that, I could just take the report from the one that I'd run and just diff them. And I was like, Oh, we got 219 out of 500, which is 43. 8%, which is to my knowledge, at least right now, state of the art also, which makes sense. But also GPT 4. 0 gets, I believe, 33%, which is like, I double checked that. The August one, the new one. Yeah, it's in their blog post. I can't remember which one it was. I don't know what the model version was. But, GPT 4, I believe, gets 33%. Which is, obviously, significantly better than what it got on the, um, original. Like, Sweebench, Sweebench, Sweebench. 2%! Yeah, yeah, yeah, exactly.…

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