The repository describes a human-reviewed golden set of review findings for each pull request, used as ground truth for comparing an agent's findings.
ReviewBench measures whether a reviewer identifies useful issues while avoiding false positives, and can break results down by severity and category. It includes a public test set of 25 tasks, an offline evaluation path, and a local LLM judge for tuning a reviewer.
Teams can submit reviewers to a public leaderboard through local validation, a self-service onboarding portal, and official test and final evaluations. The source does not establish that leaderboard scores predict review quality across all repositories, languages, team practices, or production workflows.
The repository is publicly available on GitHub under the MIT license.