CoolFace
Datasetpublic

stan4u/formally-verified-c-core-v1

Formally Verified C Core-v1 Core-v1 is a project-authored set of 64 fixed-contract C/ACSL function completion tasks for reinforcement-learning environment development and model evaluation. A model receives a complete C translation unit whose target body is replaced by a TODO. The unchanged ACSL contract and surrounding source define the problem; Frama-C WP+RTE supplies the executable reward signal. Contents 33 training tasks 15 validation tasks 16 held-out test… See the full description on the dataset page: https://huggingface.co/datasets/stan4u/formally-verified-c-core-v1.

sourceHugging Faceapache-2.0updated 14d agoView on Hugging Face
0likes34downloads
Dataset Card

Formally Verified C Core-v1

Core-v1 is a project-authored set of 64 fixed-contract C/ACSL function completion tasks for reinforcement-learning environment development and model evaluation. A model receives a complete C translation unit whose target body is replaced by a TODO. The unchanged ACSL contract and surrounding source define the problem; Frama-C WP+RTE supplies the executable reward signal.

Contents

  • —33 training tasks
  • —15 validation tasks
  • —16 held-out test tasks
  • —one reference implementation and at least one plausible wrong implementation per task
  • —record-level source hashes, license, origin, transformation, semantic family, derivation family, and review state

Semantic and derivation families are isolated to one split. This prevents a template variant from appearing in both training and evaluation.

Machine-verification evidence

Under Frama-C 33.0 (Arsenic), Why3 1.8.2, Alt-Ergo 2.6.3, and Z3 4.8.12:

  • —64/64 reference implementations proved;
  • —296/296 proof obligations discharged;
  • —84/84 runtime-error obligations discharged;
  • —zero reference solver timeouts; and
  • —64/64 deliberately wrong implementations rejected by the deterministic Frama-C WP+RTE/Qed negative pass, with zero timeouts.

The repository includes the per-task JSONL evidence and a compact release manifest. These are machine-verification results, not an independent human review and not evidence that any RL algorithm improves a model.

Schema

The primary file is tasks.jsonl. Important fields include stable_id, skeleton_c, reference_solution, negative_cases, semantic_family, derivation_family, source_repository_url, source_revision, source_content_sha256, license_spdx, and review_status.

manifest.json defines the frozen family-isolated splits and records the task file checksum.

License and provenance

Core-v1 was authored for this project and is licensed under Apache-2.0. Its deterministic generator is environments/acsl-c/scripts/build_core_v1.py in the canonical repository.

Core-v1 contains no CASP-derived C, skeleton, or reference implementation. The larger CASP corpus used during local environment engineering remains an explicit, non-bundled research adapter because the snapshot available to this project does not retain sufficient per-file origin and license metadata for public redistribution.

Intended use and limitations

Core-v1 is a compact environment seed and release test pack. It is not a comprehensive C verification benchmark, does not represent production C code, and should not be used to claim state-of-the-art model performance. The trusted computing base includes Frama-C, Why3, the selected SMT solvers, the judge runner, and the sandbox runtime.

Canonical repository: <https://github.com/stanleyngugi/formally-verified-code-rl>