Edge Case
A rare but valid sample that stresses the normal labeling rules and often exposes gaps in the guidelines.
An edge case is an input that technically falls within scope but sits at the boundary of what the guidelines anticipated — a sarcastic sentence in a sentiment task, a partially visible object in a bounding box task, a request that's borderline policy-violating rather than clearly one or the other.
Edge cases are valuable precisely because they're rare: a guideline that handles 99% of typical inputs cleanly can still fail silently on the 1% that matters most for model robustness. Well-run pipelines actively mine for edge cases (see hard negative mining) rather than waiting for them to appear organically.
The correct response to an edge case is usually escalation, not improvisation — a single annotator's best guess on a genuinely novel edge case is a coin flip dressed up as confidence.
What this means for trainers
Recognizing an edge case and flagging it correctly — rather than forcing it into an existing category — is one of the clearest signals of annotator skill that reviewers look for.
Related terms
Put this into practice
Browse open AI training roles from Alignerr, Mercor, Outlier, and more.
Browse AI training jobs