Machine Unlearning Part-2: LLM Unlearning

In this video, we dive deep into LLM Unlearning and understand how machine unlearning works for autoregressive Large Language Models. Unlike simple classifiers, LLMs store knowledge as token-by-token probability patterns, which makes forgetting copyrighted, private, or sensitive data much more complex. This tutorial explains the complete idea behind Forget Set, Retain Set, Negative Log-Likelihood (NLL), Gradient Difference, Forget-Retain training loop, and stability preservation in a very intuitive way. We also discuss why the objective is reversed on forget data, why retain loss is essential…