---
library_name: datasets
tags:
  - malicious-url-detection
  - url-pretraining
  - common-crawl
  - security
  - unsupervised
  - pretraining
  - urls
  - domain-adaptation
license: apache-2.0
dataset_info:
  features:
    - name: url
      dtype: string
    - name: year
      dtype: int64
  splits:
    - name: train
      num_examples: 68858131
---

# URL Pre-training Corpus

This is a large-scale unlabeled corpus of **~68 million** public URLs sampled from Common Crawl monthly snapshots (2020–2025). It was created for, and used in, the paper **"From Lab to Production: Malicious URL Detection on Real-World Data"** (anonymous submission) to pre-train transformer-based models on URL-specific structural and language patterns.

The corpus provides temporal and structural diversity, covering a wide range of domains, path/query structures, and IP-based URLs.

## Intended Use

- **Primary use**: Masked Language Modeling (MLM) or other self-supervised pre-training objectives for URL-focused models.
- Ideal for initializing models that will be fine-tuned on downstream URL classification tasks (e.g., malicious URL detection).
- Enables domain-specific tokenization and representation learning for structural/semantic URL patterns.

## Dataset Details

- **Total samples**: ~68 million URLs
- **Columns**:
  - `url`: Raw URL string (string)
  - `year`: Year of the Common Crawl snapshot the URL was sampled from (integer, 2020–2025)
- **Filtering**:
  - Only URLs with successful HTTP 200 responses.
  - Deduplicated.
  - Limited to ≤150 URLs per site-identifier (second-level domain or IP) to reduce bias toward popular sites.
- **Sources**: Randomly sampled shards from Common Crawl monthly crawls (2020–2025).

## How to Load

```python
from datasets import load_dataset

# Standard loading (downloads indexed data)
dataset = load_dataset("JPxxx/url-pretraining-dataset")

# For memory-efficient iteration over the full dataset
dataset = load_dataset("JPcooldev/url-pretraining-dataset", streaming=True)
for example in dataset["train"].take(5):
    print(example)
```

The dataset is provided as a single large split (`train`) due to its unsupervised nature. Use `streaming=True` to avoid downloading the entire dataset at once.

## Statistics

- Temporal diversity: URLs from 2020 to 2025.
- Structural diversity: Includes short landing-page URLs, deep paths, complex queries, and IP-based addresses.

## Limitations

- Derived from public Common Crawl data (public domain) — may contain outdated links (filtered to HTTP 200 at crawl time).
- No labels — purely for pre-training.

## Citation

If you use this dataset, please cite:

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