Ground truth refers to information known to be real or true, based on direct observation and measurement rather than inference.
In generative AI applications, ground truth is the gold standard against which responses by large language models (LLMs) are evaluated. It serves as the ideal expected result in statistical models for testing research hypotheses.
The process of gathering objective data for this purpose is called "ground truthing." In practical applications, ground truth is used to evaluate the accuracy of systems or algorithms.
For instance, in testing a stereo vision system's ability to estimate 3D positions, a laser rangefinder's more accurate measurements might serve as the ground truth. Similarly, in Bayesian spam filtering, the accuracy of the algorithm's spam/non-spam verdicts depends on the ground truth of the messages used for training.