Epoch
One full pass of the training algorithm over the entire training dataset.
An epoch is one complete pass over the whole training dataset; models train for many epochs, seeing the data repeatedly, with weights updated on each mini-batch within an epoch.
Worked example: with 1,000,000 examples and a batch size of 1,000, one epoch is 1,000 gradient steps; ten epochs is 10,000 steps over the same data seen ten times. Gotcha: too many epochs on a fixed dataset invites overfitting, while very large modern pretraining often does barely one pass over a huge corpus — epochs matter most when data is limited.