Data-adaptive priors are used as a descriptive tool to unify some of the previously proposed model selection criteria in a direct manner. An empirical Bayes criterion is constructed which penalizes for overfitting of model form as well as dimensionally.