What are training, validation, and test datasets?

Training, validation, and test datasets are separate parts of a dataset used for different stages of building and evaluating a deep learning model.

Training dataset

The training set is used to learn the model parameters, such as weights and biases. During training, the model makes predictions, calculates loss, and updates its parameters using backpropagation and an optimizer.

Validation dataset

The validation set is not used to directly update the model weights. It is used during development to:

• Compare different model versions.
• Tune hyperparameters.
• Monitor overfitting.
• Decide when to stop training.

For example, if training loss keeps decreasing while validation loss starts increasing, the model may be overfitting.

Test dataset

The test set is used after model development is largely complete to estimate how well the final model performs on unseen data.

It should not be repeatedly used for tuning decisions, because doing so can indirectly overfit the model-development process to the test set.

Why keep them separate?

Using separate datasets helps provide a more realistic estimate of generalization.

A typical flow is:

Training set → Learn parameters

Validation set → Tune and select model

Test set → Final evaluation

The exact split ratio depends on dataset size and problem requirements, so there is no universal percentage that must always be used.

For related concepts, see “How do you evaluate a deep learning model?” and “What is overfitting in deep learning?”.

In short, the training set teaches the model, the validation set guides model development, and the test set measures final performance on unseen data.