A model can continue improving on its training examples while becoming less useful on unseen ones. Early stopping gives an iterative training run a defined point at which to stop when a monitored measure no longer improves. The important design decision is which data supplies that measure.
Validation data can guide training decisions, including when to stop. A final test set has a different role: evaluating the chosen procedure after those decisions are made. Repeatedly checking the test result and adjusting the stopping rule turns the supposed final test into another source of tuning information.
Write the rule before inspecting the outcome
The Keras EarlyStopping documentation exposes settings for the monitored quantity, minimum improvement, patience and whether to restore the best weights. These controls answer separate questions. Patience tolerates a run of non-improving epochs; restoring weights determines which model state remains after stopping.
For an illustrative experiment, a team might monitor validation loss, allow a short period without sufficient improvement and restore the checkpoint with the best monitored value. The chosen period is a setting to justify for the experiment, not a universal recommendation. Confirm that the metric exists in the training logs and that the improvement direction is correct.
Keep the evaluation story intact
Record the data split, monitored metric, stopping settings, stopped epoch and retained checkpoint. A training run that ended at one epoch may retain weights from an earlier one. Without that record, “the final model” can mean different things to different team members.
Use a split appropriate to the data. Related records, repeated subjects or time-ordered observations may need more care than a random row shuffle. Preprocessing must also respect the training boundary so validation information does not leak into fitted transformations.
After choosing the full procedure, evaluate the held-out test set for its intended purpose and report the result honestly. Early stopping is one control on training behaviour; it cannot repair unrepresentative data, a misleading target or leakage elsewhere. Its value comes from a transparent rule connected to a sound evaluation design.