← All articles
Model evaluation

How to build a model evaluation dataset

Design a benchmark around realistic tasks, difficult cases, and repeatable scoring.

Map examples to real tasks

List the decisions the model will support and the inputs it will see. Sample routine cases, hard cases, and high-consequence failures. Preserve those groups in the dataset so reviewers can compare performance by task rather than relying only on an overall score.

Separate diagnosis from release decisions

Use a development evaluation set to find errors and refine prompts. Keep a held-out set for comparisons after those changes. Document data provenance, scoring rules, model settings, and expected variation so a later result can be compared on the same terms.

Build a plan around supported user tasks

For an order assistant, divide examples by goals such as checking a shipment, interpreting a return policy, and requesting a change. Include ordinary requests and cases where key information is missing. Define the success criteria for each group before collecting examples. A retrieval task may need supporting documents, while an action task needs a defined state before and after the request.

Store enough context to score each case

Start with an example ID, task type, user input, required context, and acceptable outputs. Add disallowed outcomes and a scoring rubric where appropriate. Store evidence references for factual questions and expected state transitions for tool actions. If more than one answer is acceptable, record the alternatives or the criteria that make them valid instead of forcing an exact string match.

Balance coverage with honest reporting

Hard cases can be deliberately overrepresented to reveal failures. Label them accordingly and show their results separately from a sample that resembles everyday use. Track group sizes so readers can see when a high or low rate comes from a small slice. Keep benchmark examples separate from training or prompt-tuning material, including near duplicates of the same task.

Pilot the dataset using two kinds of review

Have one reviewer verify whether cases and expected answers accurately represent the application. Have another score model outputs using only what the evaluation provides. Record questions they cannot resolve without extra context. Each unresolved issue points to missing evidence, a vague scoring rule, or a task the benchmark currently cannot measure.