Akash-Sakala/phishing-site-classification

Phishing Site URL Classification

220,000 URLs labelled phishing or benign, split 70/15/15 for training and evaluating URL-only phishing classifiers. It was used to train the BERT teacher → DistilBERT student → 4-bit student distillation pipeline.

Quick start

from datasets import load_dataset

ds = load_dataset("Akash-Sakala/phishing-site-classification")
print(ds["train"][0])   # {'text': '...', 'labels': 0 or 1}

Structure

Field	Type	Description
text	string	The URL, usually without an http(s):// scheme (e.g. quebecwind.com/)
labels	int64	0 = benign, 1 = phishing

Split	Rows	Benign (0)	Phishing (1)
train	154,000	76,873	77,127
val	33,000	16,565	16,435
test	33,000	16,562	16,438
Total	220,000

Properties

• Balanced: about 50/50 phishing vs. benign in every split, so accuracy is a meaningful metric.
• Clean splits: all 220,000 URLs are unique, so no URL is shared between splits.
• Short inputs: the median URL is 43 characters and the longest is ~2.3k characters. A max_length of 128 tokens covers ~98% of rows (64 tokens covers 93%).
• No scheme: only ~0.1% of URLs start with http. Strip the scheme from inputs at inference time to match.

Examples

text	labels
paypal.ca.8464.secure3y.mx/images/cgi.bin/ca/web.scr/?cmnd=_lg&nav=…	1 (phishing)
antares63.ru/templates/beez/http/www.paypal.com/uk/cgi-bin/webscr.php?cmd=_login-run…	1 (phishing)
jp.linkedin.com/pub/narumi-takahashi/4/577/335	0 (benign)
epodunk.com/search/counties.html	0 (benign)

Source

Derived from the Kaggle dataset Phishing Site URLs by Tarun Tiwari. That dataset has 549,346 URLs labelled Good (benign) or Bad (phishing) and is released under the Open Database License (ODbL). This dataset is shared under the same licence.

How this version was built:

1. Labels were mapped Good → 0 and Bad → 1.
2. Exactly 110,000 unique URLs per class were sampled (220,000 total), so the two classes are exactly balanced.
3. The rows were split 70 / 15 / 15 into train / val / test. Every URL appears exactly once across all splits.

Baselines

Model	Test accuracy	Test F1
BERT-base teacher	0.8971	0.8945
DistilBERT 4-layer student (KD)	0.9601	0.9595

Limitations

• The dataset has URL strings only: no page HTML, WHOIS, certificate or timing data.
• It is a static snapshot. Phishing infrastructure changes quickly, so models trained here will degrade on new campaigns.
• The sampling of benign URLs is undocumented, so performance on regional or long-tail domains is unknown.