cpu-deployable

This paper presents a reproducible, license-aware knowledge distillation method for creating efficient safety classifiers for large language models that can run on CPU hardware, achieving performance comparable to larger models while reducing false alarms.