Test set
Also called Test Set / Holdout Test
The data you touch exactly once, at the very end, to get an honest read on how the model actually does.
Think of it like
The sealed final exam — you only open it when you are truly done studying.
Example
After weeks of tuning on validation, the team runs the model on the test set a single time and reports that number.
How it actually works
The test set must stay untouched during development, because every peek lets you unconsciously tune toward it and inflate the result. Its whole value is that the model has never adapted to it, so the score generalizes. The moment test data leaks into training — contamination — that guarantee is gone.
For product teams
The one number you can actually trust when deciding whether to ship.
For engineers
A one-shot held-out evaluation split; any reuse or leakage compromises its estimate of generalization.
Related
- Validation set — The dev-time signal you tune against instead.
- Data contamination — What ruins a test set silently.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome