README.md

---
license: cc-by-nc-4.0
task_categories:
- text-classification
language:
- en
tags:
- cybersecurity
pretty_name: DeepURLBench
size_categories:
- 10M<n<100M
---

# DeepURLBench

**DeepURLBench** is a large-scale benchmark dataset for real-world URL classification, developed by Deep Instinct's research team.

## ⚠️ Warning and Usage Disclaimer

This dataset contains real-world URLs labeled as **malware**, **phishing**, or **benign**, including domains that were associated with harmful or fraudulent activity at the time of collection. **Do not attempt to visit or interact with any of the URLs in this dataset.**

This dataset is intended **solely for research and educational purposes** in cybersecurity and machine learning. We **strongly recommend using it in a read-only context**, and not resolving or querying any of the included domains or IP addresses.

Deep Instinct assumes no responsibility for misuse of the dataset.

**DeepURLBench** is a large-scale benchmark dataset for real-world URL classification, developed by Deep Instinct's research team.

## Dataset Overview

The dataset includes two subsets in Parquet format:

### 🟢 `urls_with_dns`

Contains additional DNS resolution data:

- `url`: The URL being analyzed.
- `first_seen`: The timestamp when the URL was first observed.
- `TTL` (Time to Live): DNS TTL value.
- `label`: The classification label (`malware`, `phishing`, or `benign`).
- `ip_address`: List of resolved IP addresses.

### 🔵 `urls_without_dns`

Contains only the core metadata:

- `url`: The URL being analyzed.
- `first_seen`: The timestamp when the URL was first observed.
- `label`: The classification label (`malware`, `phishing`, or `benign`).
## Important Notes on Splitting

Although Hugging Face shows each loaded file under the `"train"` split by default, this dataset **does not include predefined train/validation/test splits**. 

Instead, the intended splitting strategy is described in detail in our [paper](#citation). In brief, we recommend splitting the data **chronologically by the `first_seen` field**, so that evaluation is performed on newer, unseen URLs — simulating real-world deployment.

Each subset (`urls_with_dns` and `urls_without_dns`) is designed to be loaded independently, as shown below.


## How to Load

You can load the dataset using the Hugging Face `datasets` library:

```python
from datasets import load_dataset

ds_with_dns = load_dataset(
    "DeepInstinct/DeepURLBench",
    data_files="urls_with_dns.parquet"
)

ds_without_dns = load_dataset(
    "DeepInstinct/DeepURLBench",
    data_files="urls_without_dns.parquet"
)
```


## License

This dataset is available under the CC BY-NC 4.0 License.
## Citation

@misc{deepurlbench2025,
  author       = {Deep Instinct Research Team},
  title        = {DeepURLBench: A large-scale benchmark for URL classification},
  year         = {2025},
  howpublished = {Available at: https://huggingface.co/datasets/DeepInstinct/DeepURLBench}
}