Why training too long makes performance worse: the model memorizes training noise instead of learning general rules.

Early Stopping as regularization: continuously check a held-out validation set (80/20 split) instead of guessing epoch limits.

The three-step loop and the patience buffer that prevents stopping on a random blip before rolling back to best weights.

Training loss keeps falling while validation loss bottoms out then rises — the valley is the sweet spot before overfitting.

Halt when validation loss hits its minimum, keep the best checkpoint, and avoid overtrained end-state weights.

From epochs 8–12: best validation at epoch 10 (0.191), patience fires, halt, and automatic rollback to that checkpoint.

Prevent overfitting, train faster, spend less compute, and generalize better by choosing the best model before overfitting begins.