A feature-engineered dataset of benign and phishing URLs for machine learning and large language models evaluation

Phishing websites remain a major cybersecurity threat, yet the availability of balanced and feature-rich datasets for evaluating detection models is still limited. While machine learning (ML) and large language models (LLMs) have shown strong potential in URL-based classification, most public datasets provide raw URLs without feature engineering, making reproducibility and fair comparison across models difficult. To address this gap, we present a curated dataset of 111,660 URLs, consisting of 100,000 benign samples (label 0) and 11,660 phishing samples (label 1). Each URL entry is enriched with 22 numerical lexical and structural features (e.g., URL length, domain length, digit ratio, entropy, HTTPS usage). Additionally, three string reference columns (URL, domain, TLD) are preserved for interpretability, and one label column (0 = benign, 1 = phishing), totaling 26 columns. To demonstrate its utility, we evaluate two baseline approaches: a Random Forest (RF) classifier using handcrafted features, and a MiniLM embedding model with Logistic Regression (LR). Both achieved accuracy above 96 % and ROC AUC scores exceeding 0.99 across training, validation, and test splits. This dataset represents an important step toward building reproducible and comparable benchmarks for phishing detection, bridging traditional ML and LLM-based approaches, and supporting future research on adversarial robustness and scalable security models.