Why training too long hurts: the model stops learning general rules and starts memorizing noise.

A regularization strategy that monitors a held-out validation set and keeps the best weights before memorization.

The three-step loop plus a patience buffer so noisy blips do not abort training early.

Training loss keeps falling while validation loss valleys then rises — stop at the valley.

Epochs 8–12 with best weights at epoch 10, patience, early stop, and automatic rollback.

Prevent overfitting, faster runs, lower compute cost, and better generalization on unseen data.