Rama Ramakrishnan teaches this session of MIT's 15.773 Hands-On Deep Learning, covering how large language models built on the Generative Pretrained Transformer architecture get adapted for specific tasks. He walks through differences across GPT versions, the role and quality of training data, and instruction tuning as a method for steering base models toward useful behavior. The core of the lecture is parameter-efficient fine-tuning, the techniques that let practitioners adjust a small fraction of a model's parameters rather than retraining the whole network, making customization feasible without massive compute budgets. Running 78 minutes, the session fits into the course's broader hands-on approach to deep learning, aimed at students who already have some grounding in neural networks and want to understand the practical mechanics behind adapting modern LLMs for real applications.