What is NOT a typical ratio for splitting data into training and test sets?