Applied AI·Fine-tuning and customisation
you trained for ten passes, the training loss looked wonderful, and the model got worse at anything it had not already seen.
Epochs and overfitting
Draft summary, pending review
An epoch is one pass over the training set; too many and the model memorises examples instead of learning the pattern, getting worse on anything new. Small datasets overfit fast: often 1 to 3 epochs is right, decided by a held-out validation set.