Due to the wide availability of huge amounts of data in electronic forms, the necessity of turning such data into useful knowledge has increased. This is a proposal of learning from examples. In this paper, we propose measures to evaluate the quality of training sets used by algorithms for learning classification. Our training set assessment relies on measures provided by Rough Sets Theory. Our experimental results involved three classifiers (k-NN, C-4.5 and MLP) applied to international data bases. The new measure we propose shows good results on these test cases.