In the case of a small training set, high bias and low variance models show better performance. The sole reason is that such models do not overfit easily. In the case of a large training set, low bias and high variant models show better performance. This is possible because of their capability to reflect more complex relationships.