They heavily relay on augmentation. So for Sudoku they claim to train on 1000 puzzless, but effectively they have a set of milion because each puzzle is augmented 1000 times. So if you can not achieve similar expansion in your dataset it is very unlikely to work.