An Expanded Synthetic Conversation Dataset for Multi-Turn Smishing Detection

In three linesCOVA-X, an expanded synthetic dataset of 10,985 multi-turn conversations for detecting elder-targeted smishing scams. Longformer now outperforms XGBoost (79.71% accuracy vs 78.43%), confirming transformers require larger conversational corpora. Improved generation pipeline reduces artifacts from 67.1% to 46.5%.

Summary generated by Claude — human-verified