stindardlogic/helpfulness-safety-calibration-dpo-100k

Helpfulness-Safety Calibration DPO (100K) 100,000 DPO preference pairs for calibrating the helpfulness-safety tradeoff in language models. Each example contains a prompt, a chosen response (correct handling), and a rejected response (incorrect handling) — covering both over-refusal and under-refusal failure modes. Motivation Safety-trained models often swing between two failure modes: Over-refusal: Refusing legitimate requests because they superficially resemble… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/helpfulness-safety-calibration-dpo-100k.