Founders
How to build an evaluation dataset for AI code review tools
AI code review tools: Build an evaluation dataset around the decisions the product must support.
Build an evaluation dataset around the decisions the product must support. Include ordinary cases, difficult cases, and known failure modes with documented provenance. Keep development examples separate from the examples used to judge a change.
Start with the decision
Define the unit being evaluated and the expected result before selecting examples. Record who labeled them, what disagreement means, and how the sample relates to actual use. Revisit coverage when the product or customer workflow changes.
The illustrative scenario
This is an educational workbook. The company or fund and all numerical inputs below are fictional. They are not ADIN results, customer claims, market benchmarks, or a forecast.
The fictional service suggests review comments on pull requests for an engineering team. Maintainers remain responsible for accepting a finding and deciding whether a change is ready. Review traffic varies with release cycles.
Too many plausible but unhelpful comments can increase reviewer work. A severity label has little meaning unless the team agrees which findings require action.
An evaluation sample with a denominator
Imagine a permissioned evaluation set with 23 cases and 16 that meet a rubric defined before the test. The resulting percentage describes only this invented sample. Seventy percent is not a recommended target, a safety standard, or evidence about any real system. What matters is the rubric, sample coverage, and consequences of the remaining cases.
Too many plausible but unhelpful comments can increase reviewer work. A severity label has little meaning unless the team agrees which findings require action. Break the 7 remaining cases into failure categories before interpreting the aggregate. A recoverable wording issue and an unauthorized action may require different product behavior even when both count as a failed case. Keep reviewer disagreement and missing evidence visible.
The assumed time comparison is 28 minutes before and 19 minutes after, including the stated human step. Validate the same start and end points in a real study, then include retries, repair, and handoffs. Compare actionable findings and reviewer effort on similar changes. A tool that produces more comments is not automatically finding more consequential problems. A short average time is not enough if the result cannot be used or trusted.
| Item | Illustration |
|---|---|
| Invented evaluation cases | 23 |
| Cases meeting the fictional rubric | 16 |
| Cases requiring investigation | 7 |
| Observed proportion in this exercise | 69.6% |
| Assumed human minutes before / after | 28 / 19 |
Use this worksheet
Answer the questions with a source, date, owner, and next action. Mark an unknown as unknown. Use public or permissioned information and involve a qualified specialist when a question requires expertise beyond the team’s competence.
- What is the evaluation unit?
- Where did the examples come from?
- Who labeled the expected result?
- Which failures are underrepresented?
- What data must remain held out?
Questions specific to this scenario
- What makes a finding actionable?
- Which changes are excluded?
- Who adjudicates reviewer disagreement?
Keep a labeled sample of true defects, false alarms, dismissed suggestions, and defects missed by both workflows. Record repository and change type so aggregate numbers remain interpretable.
A concrete next test
Ask two maintainers to label the same permissioned set independently before looking at the tool's findings. Investigate disagreements about severity and actionability. A shared rubric gives the next comparison more meaning than counting comments that nobody would act on.
Interpret the result
A larger dataset is not automatically more representative. An evaluation should reveal consequential behavior in the intended workflow. If repeated tuning teaches the team the answers, add independent examples before claiming a general improvement.
Compare actionable findings and reviewer effort on similar changes. A tool that produces more comments is not automatically finding more consequential problems.
Sources and method
These primary resources provide background for the concepts. The fictional scenario, calculations, and worksheet are ADIN educational examples; the sources do not report or validate those invented results. Reviewed October 11, 2026.