MeshGraphNets training dataset

Hi,
I am new to tensorflow and I am trying to understand the dataset that is used in training the MeshGraphNets models.

In the flag_simple dataset, i see that there are 1000 trajectories and each has 400 steps, so the number of training examples will be ~400k. And the model is trained with 10M steps, does that mean it reuse these 400 samples iteratively or does each step pull one example randomly from the 400k example pools?

Thanks so much for your help!