Speaker unverified: belief

Transcript context

…What is model calibration? You have a very confused face. This is very fun. Basically, a model is said to be well calibrated if the log probability score of an answer correlates well with how correct the answer is. So you want a model which... You can basically see it as the self-confidence of the model, right? So you want a model which tells you, yes, this is true, to have high probabilities of this, if it is actually true. And this thing specifically is called calibration. And it's not that hard to measure. You could use any multi-choice evaluation set to test this. I think there are more interesting datasets to build to test that. But if we have well calibrated models, it will open the door to basically being able to have models with confidence intervals about their answers. And you would be able to say, the model is highly confident about what it's saying. Or the model is in doubt, and you could give small confidence scores. I think this would be very interesting. Yeah, there's some papers at ICLR on uncertainty as well. The quick response I'll give to that is, I think it's well known that base models are better calibrated than instruct-tuned models, right? So just the instruct-tuning just screws it up, makes it overconfident, makes it too much like a human.…

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