Nathan Lambert: evaluation

Transcript context

…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? Do you mean for training or like a deployment? Deployment. I do think that people are doing it at deployment. I think we've seen safety and other things in the RLHF pipeline. Like Lama 2 is famous for kind of having this like helpfulness and safety reward models. Deep in the Gemini report is something that Gemini has like four things, which is like helpfulness, factuality, maybe safety, maybe something else. But places like Anthropic and Chattopadhyay and Bard almost surely have a classifier after, which is like, is this text good? Is this text bad? That's not that surprising, I think, because you could use like a hundred times smaller language model and do much better at filtering than RLHF. But I do think it's still so deeply intertwined with the motivation of RLHF to be for safety that some of these categories still persist. I think that's something I'll kind of settle out, I think. I'm just wondering if it's worth collecting this data for the RLHF purpose, if you're not going to use it in any way, separate model to-…

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