Unseen (held-out) design
Beginner
A chip the learning program never saw while it was learning. Testing on one shows whether the program really learned, or just remembered.
Novice
A design kept out of a model’s training data and used only to test it. Results on unseen designs show how the model will do on the next real chip; results on new runs of the training designs flatter it.
Expert
Splitting by design, not by run or by tile, is the test that matches use. Leave-one-design-out cross-validation is common because datasets hold only a handful of designs. Also check that the technology, tools and flow match the training data, since models often have to be retrained when they change.
See also: Machine learning (ML).