You missed the “commons” part here. Sure frontier labs have internal/proprietary post-training dataset (otherwise the models we use wouldn't exist), but there's no public equivalent.