Smoothing tests based on an L$_2$ error between a truncated courier series estimator and a true function have shown good powers for a wide class of alternatives, These tests have the same form of the Neyman smooth test whose performance depends on the selected order, a basis, the farm of estimators. We construct flexible data driven Neyman smooth tests by changing a basis, combining model selection criteria and different series estimators. A simulation study shows that the generalized Neyman smooth test with the best basis provides good power for a wider class of alternatives compared with other data driven Neyman smooth tests based on a fixed form of estimator, a fixed basis and a fixed criterion.