Monitor validation data
Training loss can keep falling after generalisation weakens. Validation performance estimates behaviour on unseen examples. Stop decisions should prioritise that signal.
Mission setup
Training arena
Ready
Round 1
Epoch0 / 8080 available
Training loss—Lower is better
Validation loss—Gap —
Validation score—Validation loss
Best epoch—No checkpoint
Patience0 / 6Waiting for training
Compute used0%Budget 80 units
Model healthReadyNo trend yet
Epoch progress0%
Patience countdown100%
Interactive learning curves
Round result
Current evaluation
Complete a round to receive feedback.
Stop close to the hidden best validation epoch while conserving compute.
—
No medal
Your stop—
Optimal epochHidden
Epoch difference—
Compute saved—
Best validation—
Selected validation—
Overfitting severity—
Checkpoint restored—
Performance history
Stored locallyEarly stopping learning guide
Use patience carefully
Patience allows temporary noise without ending training immediately. Short patience saves compute but risks undertraining. Long patience tolerates noise but may waste epochs.
Set minimum improvement
Minimum improvement ignores changes too small to matter. It prevents tiny fluctuations from resetting patience. The right value depends on metric scale.
Restore best weights
The final epoch is rarely the strongest checkpoint. Restoring best weights returns the model to peak validation performance. This avoids keeping degraded late training weights.
Avoid premature stopping
A plateau may be followed by useful improvement. Warm-up epochs and suitable patience protect against early interruption. Noisy tasks usually require more tolerance.
Balance quality and compute
Early stopping is both a regularisation method and efficiency tool. Strong decisions preserve validation quality while reducing unnecessary training. Efficient models cost less to develop.