wambosec/cve-reconstruction

CVE Reconstruction

This public dataset contains the complete case assets for 375 validated CVE reconstruction cases. Each case has white-box and black-box variants, giving 750 tasks. The code-only evaluator is in the residency-environments PR. The public Prime environment also provides the code-only evaluator. The task index is data/tasks.jsonl. Each row identifies a real vulnerable and fixed release and its public Prime sandbox image. cases/ contains every case manifest and bundle: source archive, image recipe, fixture/bootstrap files, challenge and evidence, reference probe, and any case-specific verifier.

The residency environment contains its own Verifiers v1 evaluator and trusted grader. It does not depend on the separate case-generation repository. The public Prime sandbox images referenced in the manifests supply the target applications. No vision or GPU is needed. Prime authentication and ordinary usage billing are needed to allocate sandboxes and run inference; local Docker builds are also supported.

Data layout

• data/tasks.jsonl: searchable case metadata and image references.
• cases/<case-id>.json: task manifest, release pair, and image reference.
• cases/<case-id>/source.tar.gz: the read-only vulnerable source view.
• cases/<case-id>/: Dockerfile, bootstrap/fixture files, reference probe, challenge, evidence, and optional case-specific verifier.

The CVE, release, evidence, reference probe, and fixed-source information are trusted evaluation data, not solver input. Fixture access notes are exposed only where needed to log into the running application in black-box mode. For blind evaluation, use the public environment loader; it passes only the intended prompt and permitted read-only source or HTTP tools to the solver. Do not put the full dataset row or case bundle into the model prompt. Reference probes may be useful for supervised training, but they must remain separate from evaluation inputs.

The case material includes upstream source code under each project's own license. There is no single license for the entire dataset; consult each upstream project before reusing its code outside this evaluation context.