Training typically runs for several epochs, shuffling the data each time. Too few epochs can leave a model underfitted; too many can make it memorize the training data and overfit.
Very large language models often see most of their data only about once, so their training is described in tokens processed rather than epochs.