Validation Set
A dataset used during model development to compare choices and tune settings before final testing.
Which model should advance to the final test?
A gallery compares two print classifiers. Both train on the same images. One model classifies 43 of 50 validation prints correctly, the other 45 of 50. The final test prints sit in a separate folder.
Model selection
If accuracy is the agreed measure and the other constraints are met, 45/50 favors the second candidate on this validation set. Inspect the two extra successes and any important new failures.
Final assessment
The winning candidate was chosen using these 50 outcomes. Its 90% validation accuracy is a development result. Evaluate the selected pipeline on the independently held-out test prints.
Why can validation performance drift upward?
Trying many settings gives more chances to fit quirks in the validation sample. Repeated manual revisions count too. Record the search history and split by the unit that matters, such as the original print rather than individual crops.

What if the sample is small?
Use a well-designed resampling scheme on development data to study variability, then state the limits of the final independent assessment. A two-print margin alone offers little evidence about which candidate will win on a new collection.
Follow the idea
Test Set · Reserve an independent final assessment.
Model Evaluation · Match the metric to the intended use.
Examples in the AI guides
Source · Google on dividing datasets for development and testing
Return to Validation Set in the A–Z glossary