An epoch is one complete traversal of the training dataset, typically divided into smaller batches. If a dataset has 10,000 examples and a batch contains 100, a conventional epoch contains 100 training steps. The model's weights are updated as batches are processed, rather than once at the end.
Training often uses multiple epochs so the model has repeated opportunities to learn patterns. But more passes are not always better. Once the system has learned useful regularities, continued training can make it fit peculiarities of the training examples instead of generalizing to new data.
In distributed or streaming training, an epoch can be less literal: examples may be sampled with replacement, shuffled across workers or drawn from a changing mixture. Research descriptions should explain what a full pass means in the specific pipeline before comparing epoch counts.
Read epoch counts alongside dataset size, token count, batch size, learning rate and validation results. Ten epochs over a small corpus are not equivalent to ten epochs over a huge corpus, and epoch count alone does not establish model quality.