Skip to content

dataset_guide.md

Latest commit

 

History

History
65 lines (46 loc) · 2.01 KB

dataset_guide.md

File metadata and controls

65 lines (46 loc) · 2.01 KB
Raw

PolyGuard — Dataset Guide

Supported Data Sources

Source Type Labels
SARD (NIST) C/C++ vulnerable functions buffer_overflow, etc.
CVE Details Multi-language real CVEs All categories
OWASP WebGoat JS/Java web vulns sql_injection, xss
GitHub advisories Python, JS All categories
Synthetic (built-in) All languages All categories

Data Format

Each sample in data/raw/*.json must follow this schema:

{
  "code": "cursor.execute('SELECT * FROM users WHERE id = ' + uid)",
  "language": "python",
  "labels": ["sql_injection"],
  "source": "synthetic"
}
  • code: raw source code string (any length; will be truncated to max_seq_length tokens)
  • language: one of python, javascript, cpp, c, typescript
  • labels: list of label names from the taxonomy; empty list [] means safe/clean
  • source: origin of the sample (for provenance tracking)

Running the Pipeline

# 1. Place raw JSON files in data/raw/
# 2. Run preprocessing to create train/val/test splits
python -m src.data_pipeline.dataset_builder

# Or via script:
bash scripts/preprocess.sh

Outputs:

  • data/processed/train.json — 80% of samples
  • data/processed/val.json — 10% of samples
  • data/processed/test.json — 10% of samples

Synthetic Data (Demo Mode)

If data/raw/ is empty, the dataset builder automatically generates 160 synthetic samples covering all 6 vulnerability types plus clean samples. This is enough to smoke-test the pipeline end-to-end, but not sufficient for a production-quality model — real labeled data is required for that.

Class Imbalance

Vulnerability datasets are typically imbalanced (far more clean samples than vulnerable ones). Strategies implemented or planned:

  • Oversampling: repeat minority-class samples in training
  • Loss weighting: BCEWithLogitsLoss(pos_weight=...) in the trainer
  • Threshold tuning: choose per-label thresholds on the validation set to maximize F1