Skip to content
BENIBIX Learn

Ground Truth

AI📏 Evaluation & BenchmarksEvaluation Basics

Also known as Gold Labels, Gold Standard, Reference Labels

The right answers a model is checked against. People often make them by hand.

Think of it like

It is like the answer key for a quiz. You check each answer against it.

Example

Lucy wants to test a model that reads receipts. She has three people type in the totals of 500 receipts, and those totals become the ground truth.

Why it matters

Every score is only as good as the answers behind it. Bad ground truth gives a score you cannot trust.

Definition

Ground truth is the set of correct answers you check a model against. Most often, people make it by hand, such as by marking which emails are spam or drawing boxes around cars in photos. It is what turns a model's guesses into a score. If the ground truth has mistakes, the score has them too. For many tasks, getting good ground truth costs more than training the model.
The model's guess beside the answer key. Two match, and one does not.

Common mistake

Trusting labels without checking them. When two people disagree on a label, the task may be harder than it looks.

Key takeaways

  1. 1
    The answer key. What is right.
  2. 2
    Often made by hand. By people.
  3. 3
    Errors carry over. Into the score.