Well this is perfect then. We just post-process models like this after training.