It follows the same style as the training data, and academic papers, which use a lot of em dashes, are an excellent source of training data.