The correct answer that a model’s output is checked against. In a photo dataset it is the label saying ‘cat’ or ‘dog’; in a translation test it is the reference translation a person wrote. Every score a model receives depends on it, so if the supposed truth is wrong or disputed, the score is too.
Technical view
The reference labels or outputs treated as correct for training and evaluation. In practice it is rarely absolute: it is produced by annotators, instruments or existing records, each with error, and for subjective tasks (toxicity, relevance, summary quality) reasonable people disagree. Good practice records how it was produced, measures inter-annotator agreement, and treats low-agreement items separately rather than forcing a single answer. Where outputs are open-ended, a single reference understates the space of acceptable answers, which is why reference-based metrics are increasingly supplemented with rubric or model-based judging.
Related
Part of AI Terminology Index, section: Data.