i tried to follow the example on full dataset and open the dense refeeding trick every iteration. make the training 36 iterations [seems do not change much after that?], but my best result on validation set is 0.914, some gap to the reported 0.91. is there other tricks to duplicate the result? thanks!