PhreshPhish: Large-Scale Phishing Webpage Detection Benchmark

498,255 training samples and 168,060 test samples of HTML-URL pairs labeled as benign or phishing. The dataset includes 975 benchmarks with base rates ranging from 5e-4 to 5e-2 and was updated in February 2026 with approximately 200,000 new samples collected between March and December 2025.

Use Cases

• Training phishing detection models based on HTML-URL pairs.
• Benchmarking model performance across varying base rates from 5e-4 to 5e-2.
• Evaluating model robustness on real-world, large-scale web data.
• Studying temporal trends in phishing techniques based on data collected between March and December 2025.

Strengths

• Large scale with 498,255 training samples and 168,060 test samples.
• Explicitly designed for benchmarking with 975 benchmarks at varying base rates.
• Includes recent data, with an update adding ~200k samples collected in 2025.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect temporal bias inherent to the 2025 collection period.

Provenance

Source

huggingface

Collection Method

Likely collected from real-world web sources.

Time Range

Data collected between March and December 2025.

Freshness

Last updated 2026-02-09 16:38:43.