What does “Ground Truth” mean?
Trusted reference labels used to train or evaluate model outputs.
How Ground Truth Works
Ground Truth comparing Model output (Predicts: 'cat') versus Ground truth (Verified label: 'dog').
Practical Example
For a tumor-detection model, doctors confirm the diagnosis from biopsy results, and that confirmed answer is the ground truth the model's predictions are scored against.
Why It Matters
Training and evaluation are only as trustworthy as the reference answers. Mistakes in ground truth can make good models look bad, or the reverse.
Common Questions About Ground Truth
Q: Is ground truth always correct?
No. Human labels can be wrong or subjective, so teams measure agreement and review disputed cases.
Q: Where does it come from?
From human annotators, expert review, measurements, or records of what actually happened, depending on the task.
Q: Does every task have one?
Not clearly. For open-ended tasks like summarizing or writing, many answers can be good, so evaluation relies on rubrics or human judgment.