Classifier-free guidance in LLMs Safety

Roman Smirnov

The paper describes LLM unlearning without a retaining dataset, using the ORPO reinforcement learning method with inference enhanced by modified classifier-free guidance. Significa…