LLM-PHISH-900: A Hand-Crafted Dataset of LLM-Styled Phishing Emails for Transformer-Based Detection Research

LLM-PHISH-900 is a dataset of 900 hand-crafted phishing emails written to emulate the documented linguistic styles of four LLM families (Claude Sonnet, GPT-4o, Gemini 1.5 Pro, LLaMA-3 70B) across 8 social-engineering attack categories (e.g. Password Reset, CEO Fraud/BEC, Credential Harvesting, HR/Benefits Phishing). It was built to evaluate zero-shot recall and augmentation-based recovery of transformer-based phishing detectors against LLM-styled attacks. IMPORTANT: emails are hand-crafted proxies matching published stylometric patterns for each LLM, not text sampled via live API calls to those models — see README.md and DATASHEET.md in the uploaded files for full provenance details and limitations.