Customizing Open-Source LLMs

This comprehensive course guides students through the fine-tuning of open-source Large Language Models (LLMs), with a particular focus on Meta LLaMA 3. Starting with an introduction to LLM architecture and capabilities, students will learn the fundamentals of PyTorch, including tensor operations and neural network construction. These basics are essential for understanding the mechanics of LLMs.

The course emphasizes data preparation, covering tokenization, text normalization, and preprocessing techniques to ensure datasets are optimized for fine-tuning. Through hands-on exercises, students will gain practical experience in managing and preparing various types of text data.

The core of the program involves fine-tuning Meta LLaMA 3 using PyTorch, with topics such as model loading, optimization strategies, regularization techniques, and troubleshooting. Advanced concepts like cloud-native training with Nvidia NIM and deploying models using Docker and Kubernetes are integral parts of the curriculum. Students will also learn to export models for inference and build scalable inference pipelines, focusing on monitoring and maintenance for reliable performance.

The course concludes with a capstone project, where students will fine-tune and deploy Meta LLaMA 3 on a chosen platform, showcasing their mastery of the entire pipeline from data preparation to cloud-native deployment.