Test Set
A held-out portion of labeled data touched once, at the end, to estimate generalization.
- Used after model selection and tuning are complete.
- Touching it during development causes information leakage.
- Repeated test-set use is a slow form of overfitting to it.
When to use: Always; report the test metric exactly once, then resist the urge to tweak.
Example: After training and validation tuning, the model scores 87% accuracy on the test set. That is what gets reported, with no further tuning afterward.