A Sheng Phishing Corpus for Low-Resource Cybersecurity NLP

This dataset, the Sheng-English Phishing Corpus (SEPC), was created to address the critical lack of resources for phishing detection in low-resource, code-mixed languages. It contains 9,970 curated samples of legitimate (ham) and phishing messages. The data is specialized for Sheng, a dynamic Swahili-English sociolect spoken in Kenya. The corpus was constructed using a novel pipeline, including automated transcription of online video content (from YouTube, TikTok, FB and X) using OpenAI's Whisper, and augmented with data from social media, Sheng dictionaries, and web scraping. The full data collection methodology is detailed in our accompanying paper, "Hybrid Knowledge Distillation and Federated Learning for Real-Time, On-Device Phishing Detection in Low-Resource Languages." The final dataset is provided in CSV format with two columns: 'text' (the message content) and 'label' (0 for legitimate, 1 for phishing).