Glossary
Ground Truth
The accepted correct answer or label for a data example, used as the standard against which model outputs are measured.
Ground truth is the agreed-upon “right answer” for a given example — the label or output that, for training and evaluation purposes, represents what the model should produce. In some cases, ground truth is clear-cut: a factual question has a verifiable answer. In others, it’s constructed through human judgment: multiple raters review a response and their consensus becomes the ground truth. One of the recurring challenges in model behavior work is that ground truth for subjective qualities — helpfulness, tone, appropriateness — is genuinely contested. Being explicit about how ground truth is established, and what assumptions it encodes, is important for interpreting evaluation results honestly.