Wouldn't the model start overfitting at some point? Degrading generalization for accuracy on the training set.

You’d think so, but I haven’t seen it explicitly discussed in their papers, and nobody else that I know of trains on that many tokens