Harryis/LifelongSA

This is the two iteration defender of NeurIPS 2025 "Lifelong Safety Alignment for Language Models": https://openreview.net/forum?id=9YkEcAqiIK The defenders are trained on RR and LAT.