Build a dataset around real decisions
Start with the application’s intended task and harms, then collect examples that reflect real usage with appropriate privacy controls. Include typical requests, rare but consequential cases, ambiguous questions, adversarial inputs, and requests that should be declined or escalated. For a grounded assistant, retain the permitted evidence available at the time, an expected outcome where one exists, and the relevant access context. Keep development examples separate from a final holdout, and group related conversations or documents to reduce leakage.
Design six evaluation cases for a policy assistant: two ordinary questions, one ambiguous, one unanswerable, one unauthorized, and one prompt-injection attempt. For each, state what an acceptable response must do.
Separate measurable dimensions
A response can be fluent yet incorrect, or supported by a retrieved document that is itself obsolete. Score task completion, factual correctness, evidence support, safety, and required format separately. For retrieval systems, also inspect whether the relevant authorized evidence was retrieved. Deterministic checks are useful for exact constraints such as a valid schema, but word overlap alone may miss a correct paraphrase or reward an incorrect answer with similar wording.
A response quotes a retrieved policy accurately but the policy expired last week. Which dimension might pass, which might fail, and what metadata would help diagnose the issue?
Calibrate the judge against people
A model-based judge applies a rubric to candidate responses, but its scores can disagree with expert reviewers and may favor certain styles or positions. Prepare a human-rated sample spanning normal and difficult cases, define adjudication rules, and compare judge results with human labels by category and severity. Inspect false passes and false failures rather than relying only on average agreement. Keep a separate calibration holdout when refining the judge, and repeat calibration after model, rubric, or task changes.
Human reviewers identify 20 unsafe answers in a sample. A judge flags 15 of those 20 and also flags 10 safe answers. Its recall on unsafe answers is 75%; its precision among flagged answers is 15/25 = 60%. Explain why both numbers matter.
Use evaluation as a release and monitoring process
Run the same versioned cases against the baseline and candidate, with consistent retrieval and configuration or with differences explicitly recorded. Review aggregate results and important slices, uncertainty from small samples, latency, and cost. Define failure thresholds for critical cases and a human escalation path before launch. After release, sample and review production failures with consent and data controls, add verified regressions to development tests, and refresh the holdout when the workload changes.
A prompt update improves average helpfulness but fails two unauthorized-data cases. Write a release decision, identify evidence to retain, and describe how to test the fix without exposing sensitive data.
A judge flags 15 of 20 unsafe answers and also flags 10 safe answers. What are recall and precision for detecting unsafe answers?
Keep exploring
- Generative AI evaluation overview ↗
- Evaluate a judge model against human ratings ↗
- Define evaluation metrics ↗
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.