Thanks for this wonderful dataset and paper. However, I has some question of training detail. # question The hyperparameters of the code start in the paper and the repo are different. If I want to reproduce the experiment, which one should I follow?