Reference · Glossary

Benchmark

Last updated

A **standardized test suite** used to compare models on tasks like math, coding, safety, or retrieval — same questions, scored the same way.

When to use

Choosing between models, tracking regressions after prompt changes, or reading vendor leaderboards with healthy skepticism.

When not to

High benchmark scores don't guarantee your product works — real users, your docs, and your latency budget may differ completely.

Example

MMLU measures broad knowledge across subjects; a model that tops MMLU can still hallucinate your company's refund policy without RAG.