As another poster indicates, that's specifically not the case. Most possible arithmetic problems of reasonable size aren't anywhere in the dataset, but it can solve them with pretty good accuracy.