Training
Epoch
An epoch is one pass through the training dataset.
How it works
Epochs are a data-coverage unit, not a fixed amount of compute across datasets. Shuffling changes which examples share a batch on each pass.
A concrete example
Illustrative: 1.2 million examples at batch size 1,200 make 1,000 optimizer steps per epoch.
Do not confuse it with Step
a step is one parameter update; an epoch contains many steps.