PhishTrap

Catch phishing URLs before they catch you — 16 features, 19,954 URLs, balanced 50/50. Cross-verified from 496K phishing domains + Tranco top 10K. Automatically refreshed every 6 hours.

Priorities: Quality > Ease of Access > Quantity

Build pipeline (open source): github.com/instax-dutta/PhishTrap — see how every row is fetched, merged, deduplicated, validated and published.

---

Dataset Overview

PhishTrap is a curated phishing URL detection dataset built for reproducibility. Every URL is either a confirmed phishing domain (from PyFunceble-verified feeds) or a top-ranked legitimate domain (from Tranco). Features are extracted from URL structure alone — no third-party lookups, no page content.

Property	Value
Total Samples	19,954
Phishing URLs	9,977 (50%)
Legitimate URLs	9,977 (50%)
Features	16
Train/Val/Test	70/15/15
Formats	CSV, Parquet
License	MIT

---

Features

Original 6 — URL Structure

Feature	Description
url_length	Total URL character count
hyphen_count	Number of hyphens in URL
digit_count	Number of digits in URL
subdomain_count	Number of subdomains
trusted_tld	1 if TLD is .com/.org/.net/.edu/.gov
protocol_exists	1 if http/https present

Extended 10 — Deeper Signals

Feature	Description
special_char_count	Count of special characters (@, -, _, ., etc.)
entropy	Shannon entropy of URL string
path_depth	Number of path segments
domain_length	Length of domain portion
is_domain_ip	1 if domain is an IP address
has_at_symbol	1 if @ present in URL
has_double_slash_redirect	1 if // appears after protocol
tld_length	Length of top-level domain
query_param_count	Number of query parameters
path_length	Length of path portion

---

Data Sources

Phishing URLs

Source	Domains	Verification
Phishing.Database	496,576	PyFunceble-tested daily (MIT)
OpenPhish	300+ (latest feed)	15-min update cycle

Phishing domains appear in at least one verified source. Domains are deduplicated across sources.

Legitimate URLs

Source	Domains	Why
Tranco top 10K	10,000	Aggregates CrUX + Cloudflare + Umbrella + Majestic

Legitimate domains are filtered against the phishing set to remove any overlap.

---

Label

• 0 = legitimate
• 1 = phishing

---

Usage

With Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("saidutta69/PhishTrap")
train = dataset["train"]
val = dataset["val"]
test = dataset["test"]

# Train a classifier
from sklearn.ensemble import RandomForestClassifier

X_train = train.select_columns([c for c in train.column_names if c not in ("url", "label", "source")])
y_train = train["label"]
X_test = test.select_columns([c for c in test.column_names if c not in ("url", "label", "source")])
y_test = test["label"]

clf = RandomForestClassifier(n_estimators=100, random_state=42)
clf.fit(X_train, y_train)
print(f"Accuracy: {clf.score(X_test, y_test):.4f}")

With Pandas

import pandas as pd

df = pd.read_csv("hf://datasets/saidutta69/PhishTrap/data/phishtrap_full.csv")
print(df.head())

---

Auto-Update Pipeline

PhishTrap is continuously refreshed by an automated pipeline. The dataset is rebuilt every 6 hours with fresh phishing URLs from live feeds. The full pipeline is open source at github.com/instax-dutta/PhishTrap.

Architecture

GitHub Actions (cron, every 6h)
  │
  ├─ 1. Fetch phishing URLs  ── OpenPhish (15-min feed) + Phishing.Database + PhishStats
  ├─ 2. Fetch legit domains  ── Tranco top 10K
  ├─ 3. Merge with existing   ── canonical dedup (scheme/port/trailing-slash agnostic)
  ├─ 4. Prune expired         ── phishing URLs unconfirmed > 60 days are dropped
  ├─ 5. Extract 16 features   ── same feature set across all versions
  ├─ 6. Rebalance 50/50       ── stratified sampling
  ├─ 7. Split 70/15/15        ── train/val/test
  └─ 8. Validate + upload     ── pre-upload gates + post-upload verification → Hugging Face

Quality guarantees

• Freshness tracking — every row carries first_seen, last_seen, and collected_at timestamps; consumers can tell how current each URL is
• Canonical dedup — http://evil.com, https://evil.com/, and http://evil.com:80 collapse into one row via the canonical_url column
• New evidence wins — if a legitimate Tranco domain later appears in phishing feeds (e.g. it got compromised), the phishing label overrides it; label conflicts are logged
• Expiry pruning — phishing URLs unconfirmed for more than 60 days are removed automatically; legitimate domains never expire (Tranco top-10K is stable by design)
• Provenance preserved — the source column records every feed that reported a URL (e.g. openphish+phishing_database)
• Source resilience — 3 retries with exponential backoff per fetch; an unreachable source is skipped, never fatal
• Pre-upload validation — every split is checked for missing features, NaN/inf values, duplicates, and class balance before anything is uploaded
• Post-upload verification — after pushing, the pipeline downloads the uploaded files and re-validates them
• Build metadata — data/dataset_info.json records build ID, date, class/source distributions, and schema version
• No partial state — the full data/ folder is replaced atomically per run

Running the pipeline

export HF_TOKEN=hf_xxx          # required, from HF account settings
python run_pipeline.py          # incremental refresh
python run_pipeline.py --force-full    # rebuild from scratch
python run_pipeline.py --max-samples 20000   # bigger dataset
python run_pipeline.py --dry-run       # fetch/merge/prune/rebalance, no upload

Manual trigger

The GitHub Actions workflow Refresh PhishTrap Dataset can also be triggered manually (workflow_dispatch) with optional force_full and max_samples inputs.

---

Why PhishTrap?

Most phishing datasets throw URLs at you and call it a day. PhishTrap is different:

• Cross-verified sources — Every domain earns its place from reputable feeds
• No data leakage — Separate train/val/test splits with no overlap
• Clean features — 16 URL-based features, zero third-party API calls needed for inference
• Reproducible — Methodology is fully documented; re-run the pipeline anytime

---

Citation

@misc{saidutta69_2026_phishtrap,
  author = {Sai Dutta Abhishek Dash},
  title = {PhishTrap: High-Quality Phishing URL Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/saidutta69/PhishTrap}}
}

---

Built with Phishing.Database, OpenPhish, and Tranco. MIT licensed.