The aim of this work is to benchmark scoring functions used by Bayesian network learning algorithms in the context of classification. We considered both information-theoretic scores, such as LL, AIC, BIC/MDL, NML and MIT, and Bayesian scores, such as K2, BD, BDe and BDeu. We tested the scores in a classification task by learning the optimal TAN classifier with benchmark datasets. We conclude that, in general, information-theoretic scores perform better than Bayesian scores.