README.md

metadata

license: mit
task_categories:
  - text-classification
tags:
  - code
  - cybersecurity
  - vulnerability-detection
  - robustness
  - c
  - java
pretty_name: DR Evaluation Noise Pools
size_categories:
  - 10K<n<100K

DR Evaluation Noise Pools

This dataset contains the retained behavior-preserving code variants and realized-distance metadata used to evaluate the robustness of LLM-based vulnerability analyzers. The corresponding evaluation and Acc-based DR/SIR implementation is available in the TOSEM artifact repository.

Files

File	Description
noise_pools_all.zip	Combined Juliet, PrimeVul, and MegaVul perturbation pools

Archive properties:

• compressed size: 494,118,658 bytes (471.23 MiB);
• uncompressed size: 7,188,405,146 bytes (approximately 6.69 GiB);
• JSON pool files: 22,293;
• SHA-256: 2EFC57199F3E497D48772FB322D97EA7F1393BC8C1D24EB00D2E1D224F6ED824.

Evaluation Scope

The reported evaluation uses five dataset-language subsets:

Dataset	Language	Paired base files
Juliet	Java	244
Juliet	C	363
PrimeVul	C	196
MegaVul	Java	335
MegaVul	C	335

Only these dataset-language combinations are part of the reported benchmark. The archive is preserved byte-for-byte for reproducibility; consumers should use the combinations above when reproducing the evaluation.

Perturbation Families

Directory	Family	Target levels
comment_noise	Contradiction Comments (CC)	0.2, 0.4, 0.6, 0.8
prompt_inject_noise	Prompt-Injection Comments (PI)	0.2, 0.4, 0.6, 0.8
variable_noise	Variable Replacement (VR)	0.2, 0.4, 0.6, 0.8
structure_noise_checked	Structure Perturbation (SP)	0.03, 0.10, 0.17, 0.24

After extraction, the top-level layout is:

noise_pools_juliet/
noise_pools_megavul/
noise_pools_primevul/

Within each root, records are organized as:

<edit_family>/<language>/noise_<target_level>/<base_record>.json

JSON Record Structure

Each JSON pool record contains:

• cleaned_code: normalized base source code;
• formatted_code: formatted source representation;
• cwe_id: associated CWE identifier;
• one function-specific key containing retained Combo_* variants.

Each retained variant contains the edited code in code_comment_variant, safe and vulnerable harnesses in harness_command_good and harness_command_bad, and a meta_info object. In meta_info, score is the realized normalized perturbation distance consumed by the DR estimator. Comment-based families also retain code_comment_org where applicable.

Download and Extract

Download with huggingface_hub:

from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="LLMs4CodeSecurity/DR_Evaluation",
    filename="noise_pools_all.zip",
    repo_type="dataset",
)
print(archive)

Extract with Python:

from pathlib import Path
from zipfile import ZipFile

archive = Path("noise_pools_all.zip")
with ZipFile(archive) as bundle:
    bundle.extractall(archive.parent)

PowerShell:

Expand-Archive -Path .\noise_pools_all.zip -DestinationPath . -Force

Bash:

unzip noise_pools_all.zip

Evaluation Code

Clone the companion code repository:

git clone https://github.com/Jackline97/TOSEM-artifact.git
cd TOSEM-artifact
python -m pip install -r requirements.txt

For example, after placing and extracting the archive in the repository root:

python evaluation_batch.py \
  --dataset juliet \
  --language java \
  --noise-base noise_pools_juliet \
  --noise-type comment_noise \
  --models deepseek-chat \
  --noise-scales 0.2,0.4,0.6,0.8 \
  --max-combo 20 \
  --num-runs 1 \
  --out-base eval_res

See the artifact README for model configuration, Ollama usage, output semantics, and DR/SIR commands.

Intended Use and Limitations

The package is intended for research on vulnerability analysis, robustness, and behavior-preserving code transformations. It contains vulnerable code and should not be deployed as production software. The package provides retained inputs and transformation metadata; model inference outputs are generated by the companion evaluation code.

The source samples originate from Juliet, PrimeVul, and MegaVul. Users remain responsible for following the applicable terms of the upstream datasets and projects represented in those corpora.

Citation

@misc{dr_evaluation_artifact_2026,
  author       = {{LLMs4CodeSecurity}},
  title        = {DR Evaluation Noise Pools},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/LLMs4CodeSecurity/DR_Evaluation}},
  note         = {Evaluation code: \url{https://github.com/Jackline97/TOSEM-artifact}}
}