Excellent work! May I ask how the training and test sets were split for Dycheck and NVIDIA datasets during the evaluation process?