Stop near the best validation epoch
Watch the live curves, manage patience, save compute, and avoid stopping too early or too late.
Live training canvas
The marker crosses underfitting, optimal-fit, and overfitting zones as the network learns.
Performance curves
Use validation behaviour, not training improvement alone, to choose the stopping point.
Training stopped
Challenge setup
Change the scenario, difficulty, and game mode before starting.
Model and data controls
Tune the learning process and curve shape.
Coach and live signals
Hints adapt to the current curve behaviour.
Achievements
Unlock badges by stopping accurately, saving compute, and handling difficult curves.
Round history
Compare stopping decisions, validation performance, compute savings, and scores.
| # | Scenario | Mode | Stop | Best | Validation loss | Gap | Score |
|---|---|---|---|---|---|---|---|
| No completed rounds yet. | |||||||
How to make a strong stopping decision
Watch validation loss first
Training loss can keep falling after the model begins memorising. Validation loss reveals whether performance on unseen data is still improving.
Use patience for noisy curves
A single weak epoch may be temporary. Patience waits for several non-improving epochs before stopping, while minimum improvement ignores tiny changes.
Restore the best checkpoint
When training stops after the best epoch, restoring saved weights returns the model to its strongest validation state instead of using the final state.
Balance accuracy and compute
Extra epochs cost time and energy. A reliable model near the best validation epoch can be preferable to a marginally stronger but wasteful training run.