6 ms·
No, it's from RLHF. I think there are few or no examples in the training set where this kind of phrasing is referring to people. The obvious fix is to improve t
by visarga 3y ago
No, it's from RLHF. I think there are few or no examples in the training set where this kind of phrasing is referring to people. The obvious fix is to improve the RLHF dataset with more examples derived from false positives. But the real tragedy is that RLHF can make a model less creative by mode collapse.