Ground truth is the set of labels treated as correct, the reference a model is trained against and evaluated on. In annotation it is the verified, human-approved answer for each sample, the standard predictions are compared to when measuring accuracy.
Ground truth is the trusted reference: the labels you accept as correct for a dataset. A model learns by minimizing the gap between its predictions and the ground truth during training, and its accuracy is the agreement between predictions and ground truth on held-out data. Ground truth is not automatically "true," it is a human judgment about what the right answer is, which is why how you produce and verify it matters as much as the labels themselves.
Key takeaways
Ground truth is the reference labels a model trains against and is measured against.
It is a human-defined standard, not an absolute, so its quality sets a ceiling on both training and evaluation.
Disagreement among annotators is a signal that the ground truth itself is uncertain.
How ground truth is established
Define the schema or ontology so "correct" is unambiguous.
Label, then review, often with multiple annotators and a consensus or adjudication step.
Measure inter-annotator agreement to estimate how reliable the ground truth is.
Optionally keep a gold set, a small, expert-verified slice used to benchmark both annotators and models.
How it works, and how FiftyOne fits
In practice, ground truth and predictions live side by side. In FiftyOne you store the ground-truth labels and the model's predictions on the same samples, then evaluate one against the other to surface where the model is wrong, and just as often, where the ground truth itself is wrong.
Ground truth vs related terms
Definitions for ground truth, prediction, and gold set
Definitions for ground truth, prediction, and gold set
Term
What it is
Ground truth
The labels accepted as correct, the reference
Prediction
What a model outputs, scored against ground truth
Gold set
A small, expert-verified subset used as a benchmark
Why it matters
Every accuracy number is only as trustworthy as the ground truth behind it. Teams often chase model errors that are actually ground-truth errors. When a model "fails" on a sample, a meaningful share of the time the label is wrong, not the prediction. Independent audits put label-error rates at roughly 3.4% on average across common benchmark test sets, and about 6% in ImageNet's validation set (Northcutt et al., 2021). So reviewing low-agreement or high-loss samples frequently fixes the data rather than the model. Treating ground truth as fixed and infallible is one of the most common and expensive mistakes in an ML pipeline.
Frequently asked questions
Is ground truth always correct?
No. It is the agreed reference, but it can contain errors, which is why agreement and review matter.
What is the difference between ground truth and a prediction?
Ground truth is the accepted answer, a prediction is the model's guess scored against it.
How do you improve ground truth quality?
Clear guidelines, multiple annotators with consensus, agreement metrics, and a gold set.