Evaluation & Operations
Evaluation
Measuring how good a model actually is, from accuracy on tests to human judgments of quality and safety, often called evals.
Evaluation, or evals, is how you know whether a model works and whether a change helped, using a mix of automated benchmarks, task-specific tests, and human review. For generative models it is genuinely hard, since there is often no single correct answer, so teams combine scored benchmarks, head-to-head comparisons, and increasingly other models acting as graders. Good evals are the backbone of trustworthy AI development, and weak ones are a common reason products disappoint in the real world.
Related terms
Benchmark
A standardized test set used to measure and compare model performance on a specific capability.
Leaderboard
A public ranking of models by their scores on benchmarks or head-to-head human preference votes.
Hallucination
When a model states something false or made up as if it were fact, fluently and confidently.
Red Teaming
Deliberately probing an AI system to find ways it can fail, be misused, or produce harmful output, before attackers do.