What Is Validation Loss? The Number That Reveals Overfitting
Training loss tells you how well a model fits data it has already seen — validation loss is the number that tells you if it actually generalizes.
Loss measured on data the model never trained on
Validation loss is a model's loss measured on a held-out set of data it never trains on — separate from training loss, which is measured on data the model has directly learned from and is therefore a less reliable signal of real-world performance.
Why the gap between the two numbers matters most
Training loss can keep improving even after a model stops actually getting better in any meaningful sense, simply by memorizing more specifics of its training data. Validation loss reveals this — when training loss keeps dropping while validation loss stalls or rises, that gap is the clearest sign of overfitting.
How it gets used during a training run
Validation loss is typically checked at regular intervals throughout training, and is the primary signal behind early stopping and model selection — deciding when to stop training, or which saved checkpoint actually represents the best-generalizing version of the model.
Frequently Asked Questions
What is validation loss?
A model's loss measured on a held-out set of data it never trains on, providing a more reliable signal of real-world generalization than training loss, which is measured on data the model has directly learned from.
Why is the gap between training loss and validation loss important?
When training loss keeps improving while validation loss stalls or rises, that growing gap is the clearest sign of overfitting — the model is increasingly memorizing training specifics rather than learning generalizable patterns.