Tomek Korbak comments on Pretraining Language Models with Human Preferences

[Tomek Korbak]

Good question! We’re not sure. The fact that PHF scales well with dataset size might provide weak evidence that it would scale well with model size too.