Speaker unverified: recommendation

Transcript context

…We've got time, so... Basically, to go back to the initial separation, just to make it clear, so automated benchmarks, like the one we're using on the OpenLLM leaderboard, are usually fair and reproducible. Every model gets evaluated in exactly the same way, and you can really reproduce the scores you get. They tend to be also limited in the scope of what they allow to evaluate, because if you're looking at a multi-choice question, well, it's not telling you how good the model is at generating poetry, for example. So people have been using human evaluations to kind of go further in terms of the capabilities we can evaluate. We've got three types of human evaluations, in my opinion. We've got vibe-check evaluations. We've got Arena-type evaluations, like the LMsys Chatbot Arena. And then you've got human experts, so paid human annotators who will evaluate stuff, which is the approach that Scale has, for example. I think that paid human experts is a really good way to evaluate models, because you can actually give a proper grid of things you want people to check. And because they are actually paid to do so, you can hope for quite a high quality. But since human experts are expensive, people have tried to use model-as-judges, which is our third approach. Model-as-judges, I won't delve too much into this at the moment, but I think they are a problem for the field, actually. I think people should stop using LLMS judges, because they have a lot of subtle biases that they introduce in evaluation. They tend to prefer outputs from the same families. They tend to prefer first answers, which is called a position bias. They tend to prefer long and verbose answers. They struggle with evaluating models in a continuous range. So if you absolutely want to use a model-as-a-judge for your specific use case, do not use GPT-4 also because it's closed source and it will not be reproducible at all. Use a small model such as Prometheus or JudgeLM, and just use it to give you rankings, such as this option is better than this other option. Don't ask it to give you scores, because at the moment those models are not able to do this in a proper fashion. And I saw on Twitter a couple of days ago, Aidan from Cohere, who was saying that their models have a very distinct style, because they don't train with other models' outputs. They actually took the time to gather super high quality data. And other models kind of sound the same because of this. And I think for evaluation, it's going to be the exact same problem. If you choose your model based on model evaluation, you're going to make it kind of the same as all the other models. To go back on human evaluation, if we go on this distinction of VibeCheck versus R&R versus human experts, I think that VibeChecks are actually quite necessary. If you are an engineer and you want to know which model is best for your specific use case, please do a VibeCheck. You can look at a general leaderboard, like the OpenLLM leaderboard. quite necessary. If you are an engineer and you want to know which model is best for your specific use case, please do a VibeCheck. You can look at a general leaderboard, like the OpenLLM leaderboard. It will tell you which model is best in a range of tasks. And for your use case, you need to test it yourself. For the R&R or R&R-like systems in general, they are trying to rely on wisdom of the crowd approaches. But the wisdom of the crowd tends to work for quantifiable things, right? So it was initially done to try to see if a crowd could average the weight of a pig at a farmer's market in the 18th or 19th century. And it's been reproduced by asking people to estimate a number of a marble in a jar. And for anything which is like super quantifiable, it works very well. But when you're just telling people what is a good output, it's much harder to get something reproducible and experimental science is based on reproducibility, like rigorous protocols. And when using an R&R, you're not getting that. I think that an R&R is a very good sociological experiment, however. I think it's telling you a lot about the users. It's telling you a lot about what are the prompts, how people interact with models. And I also think that you can crowdsource evaluations if you have clear metrics. For example, for red teaming. You can definitely crowdsource red teaming because whether the model gave you private information or whether the model was toxic is something you can have a strict like yes or no answer, in a sense. But for anything else, it's very limited. There were a bunch of papers which were very interesting about this at ICLR this year. There was the psychophancy paper of Anthropic, where basically they showed that humans tend to prefer models which go their way and which agree with them because we want people to like us and apparently we want models to like us too and to agree with us too.…

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