JPxxx/url-benchmark-dataset

Malicious URL Detection Benchmark Dataset

This 3-million-URL benchmark dataset was introduced in the paper "From Lab to Production: Malicious URL Detection on Real-World Data" (anonymous submission). It maintains a 9:1 benign-to-malicious ratio and uses site-aware 5-fold splitting, ensuring no site-identifier (second-level domain or IP address) appears in more than one fold.

The design enables robust evaluation of model generalization to unseen threats: by preventing overlap of site-identifiers between training and evaluation sets, models cannot rely on memorizing site-specific patterns and must instead learn structural and semantic features that generalize across sites.

Intended Use

• Primary task: Malicious URL detection classification (benign vs. malicious URLs).
• Evaluation: Use site-aware folds for train/validation/test to measure generalization to unseen threats.
• Suitable for benchmarking string-based models in multi-layer detection pipelines.

Not intended for training production models deployed on single-distribution traffic (use your private data for that).

Dataset Details

• Total samples: 3,000,000 URLs
  • Benign: 2,700,000
  • Malicious: 300,000 (9:1 ratio preserved in each fold)
• Columns:
  • url: Raw URL string (string)
  • label: 'benign' (0) or 'malicious' (1) (string or int)
  • fold: Fold ID (0–4) for site-aware cross-validation (integer)
• Key properties (as defined in the paper):
  • Diversity across structures, lengths, and sources.
  • Consistent folding (similar statistics per fold).
  • Site-aware: No site-identifier (SLD or IP) appears in more than one fold.
  • Balanced length distributions between classes.
• Sources:
  • Benign: Long-lived URLs from Common Crawl (2021 & 2025) + cleaned public sources.
  • Malicious: Merged/cleaned from PhishTank, URLHaus, and existing datasets.

How to Load

from datasets import load_dataset

dataset = load_dataset("JPxxx/url-benchmark-dataset")

# Example: train on folds 1–4, test on fold 0
train_ds = dataset["train"].filter(lambda x: x["fold"] != 0)
test_ds  = dataset["train"].filter(lambda x: x["fold"] == 0)

The dataset is provided as a single split with a fold column to support flexible site-aware cross-validation.

Statistics (per fold)

• ~600,000 samples per fold
• Identical 9:1 class balance across folds
• Comparable URL length distributions and other statistical properties

Limitations

• Constructed from multiple public sources; minor distribution shifts may exist compared to any single real-world traffic stream.
• Malicious samples primarily consist of phishing and malware-delivery URLs (reflecting availability in public sources).
• Not suitable for direct training of production models.

Citation

If you use this dataset, please cite:

@article{lab2prod2026,
  title={From Lab to Production: Malicious URL Detection on Real-World Data},
  author={Anonymous},
  year={2026},
  note = {Submitted for publication}
}