CritiqueBank-11M: A Large-Scale Dataset for Training Adversarial AI Validators
Abstract
AI safety researcher focused on adversarial evaluation methods for autonomous systems. Developed MiniCrit, an adversarial critique model family trained on CritiqueBank-11M (11.7M examples), achieving 35% reduction in flawed AI outputs across 38,000+ production validations. Research spans machine learning safety, financial AI, and multi-agent systems. Open-source contributor with models and datasets released under permissive licenses.
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 6 January 2026
- Document type
- Dataset