Validation loss

Validation loss is how badly a model does on a validation set, data held back and not used for training, measured with the same scoring used during training. Its counterpart is training loss, the error on the data the model is actively learning from.

The gap between the two is the key thing to watch when training a machine learning model: while both keep falling together, the model is genuinely learning; but when training loss keeps dropping while validation loss flattens and then starts climbing, the model has begun overfitting: memorizing its practice data instead of learning patterns that carry over, often a sign it has more parameters than the data can pin down.

That turning point is the usual cue for early stopping (keeping the version with the lowest validation loss) and for adjusting things like regularization, learning rate, or dataset size. One important caveat: the validation set is only for monitoring and tuning: a separate test set is kept aside for the final, unbiased measure of how good the model really is.