Nathan Lambert: evaluation

Transcript context

…There's something later on in this talk, which is like, but it's good to mention. And this is when you're doing this preference collection, you write out a very long document of instructions to people that are collecting this data. And it's like, this is the hierarchy of what we want to prioritize. Something amount like factuality, helpfulness, honestness, harmlessness. These are all different things. Every company will rank these in different ways, provide extensive examples. It's like, if you see these two answers, you should select this one and why. And all of this stuff. And then my kind of like head scratching is like, why don't we check if the models actually do these things that we tell the data annotators to collect? But I think it's because it's hard to make that attribution. And it's hard to test if a model is honest and stuff. It would just be nice to understand the kind of causal mechanisms as a researcher or like if our goals are met. But at a simple level, what it boils down to, I have a lot more images than I need. It's like you're having a conversation with an AI, something like type GPT. You get shown two responses or more in some papers, and then you have to choose which one is better. I think something you'll hear a lot in this space is something called a Likert scale. Likert is a name. It's a name for probably some research in economics, decision theory, something. But essentially, it's a type of scale where if you have integers from like one to eight, the middle numbers will represent something close to a tie. And the smallest numbers will represent one model being way better than the other. And the biggest numbers will be like the other models better. So in the case of one to eight, if you're comparing models A to B, if you return a one, if you really liked option A, you return eight if you really like B, and then like a four or five if they were close. There's other ways to collect this data. This one's become really popular. We played with it a bit at Hugging Face. It's hard to use. Filling out this preference data is really hard. You have to read like multiple paragraphs. It's not for me. Some people really like it. I hear I'm like, I can't imagine sitting there and reading AI-generated text and like having to do that for my job. But a lot of these early papers in RLHF have good examples of what was done. The one I have here is from Anthropic's collection demo because it was from slides that I did with Anthropic. But you can look up these in the various papers. It looks like Chat2BT with two responses, and then you have an option to say which one is better. It's nothing crazy. The infrastructure is almost exactly the same, but they just log which one you think is better. I think places like Scale are also really big in this where a lot of the labeler companies will help control like who's doing how many samples. You have multiple people go over the same sample once and like what happens if there's disagreement. g in this where a lot of the labeler companies will help control like who's doing how many samples. You have multiple people go over the same sample once and like what happens if there's disagreement. I don't really think this disagreement data is used for anything, but it's good to know like what the distribution of prompts is, who's doing it, how many samples you have, controlling the workforce. All of this is very hard. A last thing to add is that a lot of these companies do collect optional metadata. I think the Anthropic example shows a rating of like how good was the prompt or the conversation from good to bad because things matter. Like there's kind of a quadrant of preference data in my mind, which is you're comparing a good answer to a good answer, which is like really interesting signal. And then there's kind of the option of you're comparing a bad answer to a bad answer, which is like you don't want to train your model on two different issues. This is like, we did this at Hugging Base and it was like, our data was like, we don't know if we can use this because a lot of it was just bad answer to bad answer because you're like rushing to try to do this real contract. And then there's also good answer to bad answer, which I think is probably pretty reasonable to include. You just prefer the good one and move on with your life. But those are very different scenarios. I think open AIs of the world are all in good answer, good answer, and have learned to eliminate everything else. But when people try to do this in open source, it's probably like what Open Assistance saw is like, there's just a lot of bad answers in your preference data. And you're like, what do I do with this? Metadata flags can help. I threw in the instruct GPT metadata. You can see how much they collect here. And like everything from the model fails to actually complete the task, hallucinations, different types of offensive or dangerous content, moral judgment, expresses opinion. Like, I don't know exactly if they're doing this now, but you can kind of see why doing RLHF at scale and prioritizing a lot of different endpoints would be hard because these are all things I'd be interested in if I was scaling up a big team to do RLHF and like what is going into the preference data. You do an experiment and you're like, okay, we're going to remove all the data where they said the model hallucinates like just that and then retrain everything. Like, what does that do? Yeah, so hallucination is big, but some of these other metadata categories, and I've seen this in a lot of papers, it's like, does it contain sexual content? Does it express a moral judgment? Does it denigrate a protected class? That kind of stuff, very binary. Should people try to adjust for this at the RLHF layer or should they put it as a pipeline where they have a classifier as a separate model that grades the model output?…

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