The Certification Map
Metrics & Generalization
6 min+50 XPStep 1/10
The split every exam asks about
All ML evaluation rests on one rule: train and test on different data.
- Training set — examples the model learns from
- Test set — held-out examples used once, at the end, to measure real performance
- Validation set (exams sometimes ask) — a middle set used to tune choices during development
If the model memorized instead of learned, the test set exposes it. That's the entire logic.
insight
Exam phrasing to expect: 'data the model has never seen.' Whenever you see that, they're describing the test set — and the concept is generalization.