What is checkpoint rollback strategy in training Keras models?

Rollback restores the best model after training instability.

Using ModelCheckpoint with restore_best_weights ensures that even if training diverges, best weights are preserved.