Only continuous ones. You would have to reformulate it as a continuous problem, such as using separate GoodGuesser examples for each discrete category, with human labels that are 1 or 0 for a category, then have GoodGuesser guess the a likelihood between 0 and 1 for each category.