Ground Truth
Definition
Ground truth is the set of correct answers that a machine learning model is trained on and measured against, established by trusted human judgment or verified records rather than by the model itself.
Every accuracy number is relative to a ground truth. If the ground truth is wrong, the model can look accurate while being wrong, or look wrong while being right. That is why establishing reliable ground truth is one of the most important steps in building a model.
For objective tasks ground truth can come from records or measurement. For judgment-based or expert tasks it comes from qualified people, often with more than one rater and a process to resolve disagreements, so the reference answers are as reliable as possible.
Ground truth is used both to train models (as labelled examples) and to evaluate them (as the answer key for a test set).
How Pathwize relates
Pathwize builds ground truth from credential-checked experts, with a sampled second rater and disagreement review, and keeps a record of who set each reference answer.
Frequently asked
The correct answers a model is trained and scored against, set by trusted human judgment or verified records.
Accuracy scores become misleading; the model may be rewarded for wrong answers or penalised for right ones.
Qualified raters, often more than one, with a process to resolve disagreements and a record of who decided each answer.
Related terms