This work demonstrates how to generate sensitive, realistic training data for abuse detection by modeling VAWG as temporally unfolding multi-turn conversations rather than isolated toxic sentences, enabling better downstream safety systems.
ConVAWG is a framework for generating synthetic multi-turn dialogues depicting Violence Against Women and Girls scenarios. Using retrieval-grounded methods, persona seeds, crime definitions, and real case reviews, it creates realistic abuse scenarios with controlled toxicity while respecting privacy constraints that prevent releasing real conversation data.