Ground Truth

Ground truth refers to the verified correct labels or annotations associated with a dataset, serving as the authoritative reference against which model predictions are compared. Although it may contain uncertainty, noise, or subjective judgment, it is treated as the accepted standard for a given task.

In supervised learning, ground truth is essential for both training and evaluation. It is commonly established through human annotation, expert review, or validated measurements.

Building high-quality ground truth often requires multiple annotators, inter-annotator agreement checks, and adjudication to resolve conflicts.