Partitioning a dataset in training and test sets