Ground truth
In one sentence Ground truth is the answer accepted as correct, against which every model prediction is scored.
Updated
Ground truth is the answer treated as definitively correct — the reference every prediction is compared against.
When you check your weight, the hospital's calibrated scale is ground truth and the bathroom scale is the model being judged. The whole comparison rests on one assumption: the reference itself is right. Nobody re-questions the hospital scale during the check-up — that is what "ground truth" means operationally. It is the thing you have agreed to stop doubting.
In practice, ground truth comes from somewhere fallible. A radiologist marks the tumour boundary. Three annotators vote on whether a review is sarcastic. A sensor logs the actual temperature. Each source has error, and annotators genuinely disagree — on some language tasks, humans agree with each other only 80-90% of the time. A model cannot be meaningfully scored above the quality of its ground truth; chasing 99% accuracy against labels that are 90% reliable is chasing noise.
Two practical habits follow. First, audit a sample of your labels by hand before trusting any metric built on them. Second, when a model and the ground truth disagree, occasionally the model is right — mislabelled examples are common in famous public datasets, including ImageNet. The term comes from cartography, where "ground truth" meant physically visiting the terrain to verify what the aerial photo suggested.
Where to go next
- Full lesson: Model evaluation
- Related terms: label, data-labelling, accuracy, benchmark