Evaluation
Benchmark
A benchmark is a specified dataset, protocol, and metric used for comparison.
How it works
Meaningful comparisons require the same version, prompt policy, scoring rules, and contamination controls. A score without its protocol is incomplete.
A concrete example
Illustrative: evaluate 2,000 fixed items with exact-match scoring and report 1,540 correct.
Do not confuse it with Evaluation
a benchmark is a standardized evaluation instance; evaluation is the broader measurement process.