Out-of-Distribution Testing
Out-of-distribution (OOD) testing checks how a model behaves on inputs that differ meaningfully from its training or evaluation data — an unfamiliar language, an unusual input format, a topic it was never exposed to — since accuracy on familiar data says nothing about behavior once real-world input drifts outside that range.
Most reported model accuracy numbers come from a test set drawn from the same distribution as training data, which makes them a poor predictor of production behavior, because real users inevitably send inputs the model has never effectively seen — a support bot trained mostly on English tickets suddenly getting a Portuguese one, a document classifier meeting a file format that wasn't in its training corpus.
The dangerous part isn't that the model fails on OOD input — some failure is expected — it's that many models fail silently, producing a confident, plausible-looking answer instead of recognizing the input is unfamiliar and flagging uncertainty. Good OOD handling means the model (or a wrapper around it) can detect "this doesn't look like anything I know" and respond accordingly, rather than guessing with false confidence.
Building an OOD test set usually means deliberately sourcing inputs that fall outside the known training/eval distribution — different domains, languages, formats, or edge-of-scope topics — and checking not just correctness but whether the model's confidence appropriately drops.
