Datasets:
Tasks:
Text Classification
Formats:
parquet
Sub-tasks:
multi-class-classification
Size:
10K - 100K
Tags:
Synthetic
code
vulnerability-detection
software-security
formal-verification
bounded-model-checking
License:
Slim release
Browse files- DATASHEET.md +3 -3
- NOTICE +2 -2
- README.md +3 -3
- croissant.json +3 -3
- data/comparison/README.md +0 -13
- data/comparison/arvo/README.md +0 -10
- data/comparison/formai/README.md +0 -9
- data/comparison/icvul/README.md +0 -9
- data/comparison/mono/README.md +0 -8
- data/comparison/rebuild_withheld_samples.py +0 -76
- data/comparison/titanvul/README.md +0 -8
DATASHEET.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# Datasheet for VeriVul
|
| 2 |
|
| 3 |
-
Follows Datasheets for Datasets (Gebru et al.). Numbers are those of the paper (NeurIPS 2026 Evaluations and Datasets Track)
|
| 4 |
|
| 5 |
## Motivation
|
| 6 |
|
|
@@ -31,7 +31,7 @@ The release also holds the samples of the five corpora that the paper compares V
|
|
| 31 |
| TitanVul (Li et al., https://github.com/yikun-li/TitanVul-BenchVul) | 400 | none found | identifiers and hashes |
|
| 32 |
| Mono (Gao et al., https://github.com/vul337/mono) | 400 | no LICENSE file | identifiers and hashes |
|
| 33 |
|
| 34 |
-
The evaluation pairs are up to 50 per CWE, drawn with seed 42, and each comparison with VeriVul uses the CWEs that both share. For TitanVul and Mono
|
| 35 |
|
| 36 |
## Collection process
|
| 37 |
|
|
@@ -73,7 +73,7 @@ The dataset contains vulnerable programs and inputs that trigger sanitizer repor
|
|
| 73 |
|
| 74 |
## Distribution
|
| 75 |
|
| 76 |
-
Zenodo, DOI 10.5281/zenodo.
|
| 77 |
|
| 78 |
## Maintenance
|
| 79 |
|
|
|
|
| 1 |
# Datasheet for VeriVul
|
| 2 |
|
| 3 |
+
Follows Datasheets for Datasets (Gebru et al.). Numbers are those of the paper (NeurIPS 2026 Evaluations and Datasets Track).
|
| 4 |
|
| 5 |
## Motivation
|
| 6 |
|
|
|
|
| 31 |
| TitanVul (Li et al., https://github.com/yikun-li/TitanVul-BenchVul) | 400 | none found | identifiers and hashes |
|
| 32 |
| Mono (Gao et al., https://github.com/vul337/mono) | 400 | no LICENSE file | identifiers and hashes |
|
| 33 |
|
| 34 |
+
The evaluation pairs are up to 50 per CWE, drawn with seed 42, and each comparison with VeriVul uses the CWEs that both share. For TitanVul and Mono only identifiers and SHA-256 hashes are shipped.
|
| 35 |
|
| 36 |
## Collection process
|
| 37 |
|
|
|
|
| 73 |
|
| 74 |
## Distribution
|
| 75 |
|
| 76 |
+
Zenodo, DOI 10.5281/zenodo.24017329 (https://doi.org/10.5281/zenodo.24017329), with a Hugging Face mirror (https://huggingface.co/datasets/qcri/verivul-dataset) and the framework repository (https://github.com/qcri/verivul-framework). Data under CC BY 4.0 and code under MIT. A Croissant 1.1 file with core and Responsible AI fields is `croissant.json`. It was validated with `mlcroissant`. Checksums are in `SHA256SUMS`.
|
| 77 |
|
| 78 |
## Maintenance
|
| 79 |
|
NOTICE
CHANGED
|
@@ -1,7 +1,7 @@
|
|
| 1 |
VeriVul dataset: licensing and attribution
|
| 2 |
|
| 3 |
Data (data/, croissant.json, DATASHEET.md, docs/): Creative Commons Attribution 4.0 International, see LICENSE-DATA.
|
| 4 |
-
|
| 5 |
|
| 6 |
Attribution
|
| 7 |
* The programs of data/verivul.jsonl were generated from seed functions of ReposVul and ICVul, which are released under the MIT License.
|
|
@@ -12,5 +12,5 @@ Attribution
|
|
| 12 |
- ICVul: MIT License (licenses/ICVul-LICENSE).
|
| 13 |
- FormAI: Tihanyi et al., "The FormAI Dataset: Generative AI in Software Security through the Lens of Formal Verification", https://github.com/FormAI-Dataset/FormAI-dataset. MIT License (licenses/FormAI-LICENSE).
|
| 14 |
- ARVO: Mei et al., "ARVO: Atlas of Reproducible Vulnerabilities for Open Source Software", https://arxiv.org/abs/2408.02153, https://github.com/n132/ARVO-Meta. BSD 2-Clause License (licenses/ARVO-LICENSE). The functions come from open-source projects that keep their own licenses.
|
| 15 |
-
TitanVul (https://github.com/yikun-li/TitanVul-BenchVul) and Mono (https://github.com/vul337/mono) are not redistributed because no license was found. data/comparison/ holds only identifiers and SHA-256 hashes for them
|
| 16 |
* ESBMC 8.4 verified the programs and keeps its own license.
|
|
|
|
| 1 |
VeriVul dataset: licensing and attribution
|
| 2 |
|
| 3 |
Data (data/, croissant.json, DATASHEET.md, docs/): Creative Commons Attribution 4.0 International, see LICENSE-DATA.
|
| 4 |
+
Code (framework/): MIT License, see LICENSE.
|
| 5 |
|
| 6 |
Attribution
|
| 7 |
* The programs of data/verivul.jsonl were generated from seed functions of ReposVul and ICVul, which are released under the MIT License.
|
|
|
|
| 12 |
- ICVul: MIT License (licenses/ICVul-LICENSE).
|
| 13 |
- FormAI: Tihanyi et al., "The FormAI Dataset: Generative AI in Software Security through the Lens of Formal Verification", https://github.com/FormAI-Dataset/FormAI-dataset. MIT License (licenses/FormAI-LICENSE).
|
| 14 |
- ARVO: Mei et al., "ARVO: Atlas of Reproducible Vulnerabilities for Open Source Software", https://arxiv.org/abs/2408.02153, https://github.com/n132/ARVO-Meta. BSD 2-Clause License (licenses/ARVO-LICENSE). The functions come from open-source projects that keep their own licenses.
|
| 15 |
+
TitanVul (https://github.com/yikun-li/TitanVul-BenchVul) and Mono (https://github.com/vul337/mono) are not redistributed because no license was found. data/comparison/ holds only identifiers and SHA-256 hashes for them.
|
| 16 |
* ESBMC 8.4 verified the programs and keeps its own license.
|
README.md
CHANGED
|
@@ -113,11 +113,11 @@ print(log["id"], log["esbmc_log"][-200:])
|
|
| 113 |
|
| 114 |
## Comparison corpora
|
| 115 |
|
| 116 |
-
`data/comparison/` has the evaluation pairs of ICVul, FormAI, ARVO, TitanVul and Mono that the paper compares with. ICVul (MIT), FormAI (MIT) and ARVO (BSD 2-Clause) are shipped with code and keep their licenses. TitanVul and Mono are not redistributed because no license was found, so their folders hold identifiers and SHA-256 hashes.
|
| 117 |
|
| 118 |
## More material
|
| 119 |
|
| 120 |
-
The Zenodo record has the same data: https://doi.org/10.5281/zenodo.
|
| 121 |
|
| 122 |
## License and citation
|
| 123 |
|
|
@@ -129,6 +129,6 @@ Data under CC BY 4.0. Seeds derive from ReposVul and ICVul, both MIT. The compar
|
|
| 129 |
author = {Lekssays, Ahmed and Mouhcine, Hamza and Khalil, Issa},
|
| 130 |
booktitle = {Advances in Neural Information Processing Systems, Evaluations and Datasets Track},
|
| 131 |
year = {2026},
|
| 132 |
-
note = {Dataset DOI: 10.5281/zenodo.
|
| 133 |
}
|
| 134 |
```
|
|
|
|
| 113 |
|
| 114 |
## Comparison corpora
|
| 115 |
|
| 116 |
+
`data/comparison/` has the evaluation pairs of ICVul, FormAI, ARVO, TitanVul and Mono that the paper compares with. ICVul (MIT), FormAI (MIT) and ARVO (BSD 2-Clause) are shipped with code and keep their licenses. TitanVul and Mono are not redistributed because no license was found, so their folders hold identifiers and SHA-256 hashes.
|
| 117 |
|
| 118 |
## More material
|
| 119 |
|
| 120 |
+
The Zenodo record has the same data: https://doi.org/10.5281/zenodo.24017329. The framework repository is https://github.com/qcri/verivul-framework.
|
| 121 |
|
| 122 |
## License and citation
|
| 123 |
|
|
|
|
| 129 |
author = {Lekssays, Ahmed and Mouhcine, Hamza and Khalil, Issa},
|
| 130 |
booktitle = {Advances in Neural Information Processing Systems, Evaluations and Datasets Track},
|
| 131 |
year = {2026},
|
| 132 |
+
note = {Dataset DOI: 10.5281/zenodo.24017329, code: https://github.com/qcri/verivul-framework}
|
| 133 |
}
|
| 134 |
```
|
croissant.json
CHANGED
|
@@ -61,13 +61,13 @@
|
|
| 61 |
"sdLicense": "https://creativecommons.org/licenses/by/4.0/",
|
| 62 |
"url": "https://huggingface.co/datasets/qcri/verivul-dataset",
|
| 63 |
"sameAs": [
|
| 64 |
-
"https://doi.org/10.5281/zenodo.
|
| 65 |
"https://github.com/qcri/verivul-framework"
|
| 66 |
],
|
| 67 |
"version": "1.0.0",
|
| 68 |
"dateCreated": "2026-10-11",
|
| 69 |
-
"dateModified": "2026-11-
|
| 70 |
-
"datePublished": "2026-11-
|
| 71 |
"keywords": [
|
| 72 |
"vulnerability detection",
|
| 73 |
"software security",
|
|
|
|
| 61 |
"sdLicense": "https://creativecommons.org/licenses/by/4.0/",
|
| 62 |
"url": "https://huggingface.co/datasets/qcri/verivul-dataset",
|
| 63 |
"sameAs": [
|
| 64 |
+
"https://doi.org/10.5281/zenodo.24017329",
|
| 65 |
"https://github.com/qcri/verivul-framework"
|
| 66 |
],
|
| 67 |
"version": "1.0.0",
|
| 68 |
"dateCreated": "2026-10-11",
|
| 69 |
+
"dateModified": "2026-11-10",
|
| 70 |
+
"datePublished": "2026-11-10",
|
| 71 |
"keywords": [
|
| 72 |
"vulnerability detection",
|
| 73 |
"software security",
|
data/comparison/README.md
DELETED
|
@@ -1,13 +0,0 @@
|
|
| 1 |
-
# Comparison corpora
|
| 2 |
-
|
| 3 |
-
The evaluation pairs of the five corpora that VeriVul is compared with in the paper. Each comparison uses the CWEs that the corpus shares with `../evaluation/seed_matched_400.jsonl`.
|
| 4 |
-
|
| 5 |
-
| Corpus | License | Shipped |
|
| 6 |
-
|---|---|---|
|
| 7 |
-
| ICVul | MIT | pairs with code |
|
| 8 |
-
| FormAI | MIT (repository) | pairs with code |
|
| 9 |
-
| ARVO | BSD 2-Clause | pairs with code |
|
| 10 |
-
| TitanVul | none found | identifiers and hashes |
|
| 11 |
-
| Mono | no LICENSE file | identifiers and hashes |
|
| 12 |
-
|
| 13 |
-
Each folder has a README with the source and the license. The license texts are in `../../licenses/`. `rebuild_withheld_samples.py` rebuilds the TitanVul and Mono pairs from the original corpora.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
data/comparison/arvo/README.md
DELETED
|
@@ -1,10 +0,0 @@
|
|
| 1 |
-
# ARVO samples
|
| 2 |
-
|
| 3 |
-
The ARVO evaluation pairs used in the paper's comparison with VeriVul. They are redistributed under the license of the source.
|
| 4 |
-
|
| 5 |
-
* Source: Mei et al., ARVO: Atlas of Reproducible Vulnerabilities for Open Source Software, https://arxiv.org/abs/2408.02153
|
| 6 |
-
* Repository: https://github.com/n132/ARVO-Meta
|
| 7 |
-
* License: BSD 2-Clause License, `licenses/ARVO-LICENSE` (the upstream LICENSE file)
|
| 8 |
-
* Note: The functions come from open-source projects that OSS-Fuzz tests. Each project keeps its own license and `project` and `patch_url` in each record identify it.
|
| 9 |
-
|
| 10 |
-
`eval_samples.jsonl` has 333 vulnerable and fixed function pairs from six CWEs (125, 120, 416, 476, 119, 704), fewer than 50 for some, with code, CWE and the original identifier (`pair_id`). Please cite the source when you use these files.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
data/comparison/formai/README.md
DELETED
|
@@ -1,9 +0,0 @@
|
|
| 1 |
-
# FormAI samples
|
| 2 |
-
|
| 3 |
-
The FormAI evaluation pairs used in the paper's comparison with VeriVul. They are redistributed under the license of the source.
|
| 4 |
-
|
| 5 |
-
* Source: Tihanyi et al., The FormAI Dataset: Generative AI in Software Security through the Lens of Formal Verification, https://zenodo.org/records/10725896
|
| 6 |
-
* Repository: https://github.com/FormAI-Dataset/FormAI-dataset
|
| 7 |
-
* License: MIT License, `licenses/FormAI-LICENSE` (the upstream LICENSE file). The Zenodo record of the first release lists CC BY 4.0
|
| 8 |
-
|
| 9 |
-
`eval_samples.jsonl` has 400 pairs, 50 for each of eight CWEs (120, 125, 134, 190, 191, 401, 416, 476). The programs of a pair are not related. The safe program is drawn at random from the safe pool, with code, CWE and the original identifier (`pair_id`). Please cite the source when you use these files.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
data/comparison/icvul/README.md
DELETED
|
@@ -1,9 +0,0 @@
|
|
| 1 |
-
# ICVul samples
|
| 2 |
-
|
| 3 |
-
The ICVul evaluation pairs used in the paper's comparison with VeriVul. They are redistributed under the license of the source.
|
| 4 |
-
|
| 5 |
-
* Source: Lu et al., ICVul: A Well-labeled C/C++ Vulnerability Dataset with Comprehensive Metadata and VCCs, https://arxiv.org/abs/2505.08503
|
| 6 |
-
* Repository: https://github.com/Chaomeng-Lu/ICVul
|
| 7 |
-
* License: MIT License, `licenses/ICVul-LICENSE` (the upstream LICENSE file)
|
| 8 |
-
|
| 9 |
-
`eval_samples.jsonl` has 350 vulnerable and fixed function pairs, 50 for each of seven CWEs (119, 120, 125, 190, 20, 416, 476), with code, CWE and the original identifier (`pair_id`). Please cite the source when you use these files.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
data/comparison/mono/README.md
DELETED
|
@@ -1,8 +0,0 @@
|
|
| 1 |
-
# Mono samples (withheld)
|
| 2 |
-
|
| 3 |
-
The code of the Mono samples is not shipped. The repository has no LICENSE file and its README only shows an MIT badge.
|
| 4 |
-
|
| 5 |
-
* Source: Gao et al., Mono: Is Your Clean Vulnerability Dataset Really Solvable?, https://arxiv.org/abs/2506.03651
|
| 6 |
-
* Repository: https://github.com/vul337/mono
|
| 7 |
-
|
| 8 |
-
`eval_sample_ids.jsonl` has the 400 evaluation pairs with `pair_id`, CWE, CVE, file name and the SHA-256 of the vulnerable and the fixed code. To rebuild the pairs run `python ../rebuild_withheld_samples.py --external <dir> --out rebuilt`, where `<dir>` holds `monoVul/*.json (the CVE metadata files of the Mono release)`. The script checks every pair against the hashes.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
data/comparison/rebuild_withheld_samples.py
DELETED
|
@@ -1,76 +0,0 @@
|
|
| 1 |
-
"""Rebuilds the TitanVul and Mono evaluation pairs from the original corpora and checks them against the shipped hashes.
|
| 2 |
-
|
| 3 |
-
python rebuild_withheld_samples.py --external /path/to/corpora --out rebuilt
|
| 4 |
-
|
| 5 |
-
--external holds TitanVul/TitanVul.csv and monoVul/*.json. The script writes <Corpus>_eval_samples.jsonl and prints
|
| 6 |
-
how many pairs match the SHA-256 hashes in titanvul/eval_sample_ids.jsonl and mono/eval_sample_ids.jsonl.
|
| 7 |
-
"""
|
| 8 |
-
import argparse
|
| 9 |
-
import glob
|
| 10 |
-
import hashlib
|
| 11 |
-
import json
|
| 12 |
-
from pathlib import Path
|
| 13 |
-
|
| 14 |
-
import pandas as pd
|
| 15 |
-
|
| 16 |
-
HERE = Path(__file__).resolve().parent
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
def norm(value):
|
| 20 |
-
value = str(value).strip() if value is not None else ""
|
| 21 |
-
return value if value and value.lower() not in ("n/a", "none", "null", "nan") else None
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
def sha(text):
|
| 25 |
-
return hashlib.sha256((text or "").encode()).hexdigest()
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
def titanvul(external, ids):
|
| 29 |
-
df = pd.read_csv(Path(external) / "TitanVul" / "TitanVul.csv")
|
| 30 |
-
out = {}
|
| 31 |
-
for pid in ids:
|
| 32 |
-
row = df.iloc[int(pid.split("_")[1])]
|
| 33 |
-
out[pid] = (norm(row["func_before"]), norm(row["func_after"]))
|
| 34 |
-
return out
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
def mono(external, ids):
|
| 38 |
-
wanted = set(ids)
|
| 39 |
-
out = {}
|
| 40 |
-
for path in sorted(glob.glob(str(Path(external) / "monoVul" / "*.json"))):
|
| 41 |
-
try:
|
| 42 |
-
data = json.load(open(path, encoding="utf-8"))
|
| 43 |
-
except Exception:
|
| 44 |
-
continue
|
| 45 |
-
cve = norm(data.get("cve_id"))
|
| 46 |
-
for commit_hash, commit in data.get("commits", {}).items():
|
| 47 |
-
for idx, patch in enumerate(commit.get("patches", [])):
|
| 48 |
-
pid = f"Mono_{cve}_{commit_hash[:8]}_{idx}"
|
| 49 |
-
if pid in wanted:
|
| 50 |
-
out[pid] = (norm(patch.get("func_before")), norm(patch.get("func_after")))
|
| 51 |
-
return out
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
def main():
|
| 55 |
-
ap = argparse.ArgumentParser()
|
| 56 |
-
ap.add_argument("--external", required=True)
|
| 57 |
-
ap.add_argument("--out", required=True)
|
| 58 |
-
args = ap.parse_args()
|
| 59 |
-
out = Path(args.out)
|
| 60 |
-
out.mkdir(parents=True, exist_ok=True)
|
| 61 |
-
for name, folder, finder in (("TitanVul", "titanvul", titanvul), ("Mono", "mono", mono)):
|
| 62 |
-
manifest = [json.loads(line) for line in open(HERE / folder / "eval_sample_ids.jsonl")]
|
| 63 |
-
found = finder(args.external, [m["pair_id"] for m in manifest])
|
| 64 |
-
ok = 0
|
| 65 |
-
with open(out / f"{name}_eval_samples.jsonl", "w") as f:
|
| 66 |
-
for m in manifest:
|
| 67 |
-
before, after = found.get(m["pair_id"], (None, None))
|
| 68 |
-
good = sha(before) == m["vulnerable_sha256"] and sha(after) == m["patched_sha256"]
|
| 69 |
-
ok += good
|
| 70 |
-
keep = {k: v for k, v in m.items() if not k.endswith("sha256")}
|
| 71 |
-
f.write(json.dumps({**keep, "vulnerable_code": before, "patched_code": after, "hash_match": good}) + "\n")
|
| 72 |
-
print(f"{name}: {ok} of {len(manifest)} pairs match the shipped hashes")
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
if __name__ == "__main__":
|
| 76 |
-
main()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
data/comparison/titanvul/README.md
DELETED
|
@@ -1,8 +0,0 @@
|
|
| 1 |
-
# TitanVul samples (withheld)
|
| 2 |
-
|
| 3 |
-
The code of the TitanVul samples is not shipped. Neither the repository nor the dataset card states a license.
|
| 4 |
-
|
| 5 |
-
* Source: Li et al., Out of Distribution, Out of Luck: How Well Can LLMs Trained on Vulnerability Datasets Detect Top 25 CWE Weaknesses?, https://arxiv.org/abs/2507.21817
|
| 6 |
-
* Repository: https://github.com/yikun-li/TitanVul-BenchVul and https://huggingface.co/datasets/yikun-li/TitanVul
|
| 7 |
-
|
| 8 |
-
`eval_sample_ids.jsonl` has the 400 evaluation pairs with `pair_id`, CWE, CVE, file name and the SHA-256 of the vulnerable and the fixed code. To rebuild the pairs run `python ../rebuild_withheld_samples.py --external <dir> --out rebuilt`, where `<dir>` holds `TitanVul/TitanVul.csv`. The script checks every pair against the hashes.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|