The recommended split of documents for training and evaluation in machine learning typically follows the Pareto principle, which suggests an 80-20 split. Given a total of 15 documents, 80% of 15 is 12 for training and 20% of 15 is 3 for evaluation. This usually ensures that the model has enough data to learn effectively while still having sufficient data for a robust evaluation. Therefore, the correct split would be 12 documents for training and 3 for evaluating the model.