Evaluation
Testing what a model can do and how it behaves, with benchmarks, human ratings and targeted test sets.
Also: evals
Evaluation covers capability benchmarks, comparisons by human raters, safety tests for dangerous knowledge and misuse, checks of bias and fairness, and the private test sets companies build for their own uses. No single number captures a model's quality, and results depend on prompting, the number of examples and how answers are scored. Independent groups such as Stanford's HELM project, Epoch AI and government AI safety institutes run their own evaluations, and labs publish results in model and system cards.