Glossary entry

Evaluation

The systematic measurement of a model's capabilities, quality, and failure modes.

Definition

What it means

Evaluation includes automated tests, human review, model-based grading, red teaming, and real-world monitoring. It can be performed before release and continuously after deployment.

Why it matters

Role in the system

Training only helps when teams can tell whether the model actually improved and where it still fails.

Example

In practice

Experts use a rubric to grade the factual accuracy and completeness of a financial analysis.