Could you share the checkpoint files and the dataset splits? Did you train using 1 epoch only? I would like to further train it with more epochs.