Getting Language Models to Open Up on ‘Risky’ Subjects

A new 'FalseReject' dataset addresses language models' tendency to refuse harmless prompts that sound risky. Developed by Dartmouth College and Amazon researchers, it contains 16,000 benign prompts across 44 safety-related categories, with human-annotated test and training sets. This dataset aims to retrain models to respond more intelligently to sensitive topics without compromising safety, potentially improving their real-world usefulness while maintaining appropriate caution.