Aren't they usually built on most of the same training data?