Early stopping is a simple, widely used training technique: monitor the model's performance on a validation set it never trains on, and stop training when that performance stops improving, even if the schedule called for more. Training loss on the data the model sees almost always keeps falling, but past a certain point the model is memorizing its training examples rather than learning patterns that generalize. Validation performance flattening or worsening is the signal that this overfitting has begun.
The mechanics are straightforward. Training pauses at intervals to evaluate on the validation set, keeps a snapshot of the best-performing weights so far, and stops after a set number of evaluations without improvement, a threshold called patience. When training ends, the best snapshot is restored rather than the final one. The benefits are twofold: better generalization, because the model is captured at its peak rather than after it began memorizing, and lower cost, because compute is not spent on epochs that make the model worse. In fine-tuning language models, where datasets are often small and overfitting arrives fast, early stopping is frequently the difference between a model that helps and one that parrots its training set.
At arosplatforms every training and fine-tuning run we execute for clients uses early stopping against a held-out validation set drawn from realistic usage, not from the training data. It is part of a broader discipline: define the evaluation before training starts, let the metrics decide when to stop, and never judge a model by its performance on data it has already seen.