Evaluation · 4 min read
Better benchmarks start with better questions.
Why useful evaluation starts with the decisions a model needs to get right.
Define the decision, then the metric
A high aggregate score can hide a failure that matters to a real user. Start by describing the intended task, the consequences of an error, and the conditions in which the system will operate. Those decisions help determine which examples belong in an evaluation set and which outcomes deserve closer review.
Make expert judgment reviewable
A rubric should make disagreements visible. Ask reviewers to identify evidence, distinguish factual errors from reasonable alternatives, and record uncertainty. Calibrate on shared examples before scaling. When reviewers disagree, examine the instruction and the task before assuming one person is wrong.
Look beyond the average
Inspect performance across domains, languages, difficulty levels, and failure types. Keep a holdout set separate from development feedback, and revisit assumptions when a model or deployment context changes. A benchmark is a tool for learning about behavior, not a permanent guarantee of quality.