Ground Truth

Ground truth is the authoritative, human-verified correct answer or label used as the reference point in an evaluation. A model's output is compared against ground truth to compute objective metrics such as exact match, F1, or semantic similarity. The quality of ground truth directly limits evaluation reliability: noisy or inconsistently labeled references produce misleading metrics even when the underlying model is performing well.