Definition
The training set teaches the model. The validation set guides hyperparameter choices. The test set estimates real-world performance once — after decisions are locked.
Leakage between splits makes published numbers look better than reality.
In simple terms
Study materials (train), practice exams you may peek at while studying strategy (validation), and the sealed final exam (test).
Where you see it
- 80/10/10 splits on classification tasks.
- Time-based splits for forecasting so future never leaks into past.
How it works
1.Partition carefully
Random, stratified, or temporal splits.
2.Train and tune
Use train + validation only.
3.Final report
Score once on test; avoid peeking repeatedly.
Why it matters
- Honest splits are the difference between a demo and a deployable model.
Often confused
It is fine to tune on the test set if you are careful.
Repeated test peeking turns the test set into validation and inflates scores.