11-667: Large Language Models Methods and Applications - Carnegie Mellon University

Delivers a comprehensive, sequenced curriculum covering large language model theory, architecture, and applications through structured course modules. The system organizes learning materials into progressive difficulty levels (beginner to advanced) with integrated lectures, assignments, and practical exercises that build foundational understanding of transformer architectures, attention mechanisms, training methodologies, and deployment patterns. This is implemented as a university-level course structure with curated content pathways rather than ad-hoc documentation.

Teaches the complete transformer architecture including self-attention mechanisms, multi-head attention, positional encoding, feed-forward networks, and layer normalization through mathematical derivations and conceptual explanations. The curriculum covers how attention computes query-key-value projections, why positional encoding is necessary, and how transformer stacks compose these components into a complete model. This goes beyond high-level descriptions to explain the 'why' behind architectural choices and mathematical properties.

Teaches architectural patterns for building production LLM applications, covering system design considerations, integration with existing systems, scalability patterns, and operational concerns. The curriculum covers different application architectures (simple prompting, RAG, agents, multi-model systems), how to structure applications for reliability and maintainability, and how to integrate LLMs with databases, APIs, and other services. This includes both high-level architectural patterns and practical implementation considerations.

Teaches practical and theoretical aspects of training large language models from scratch and fine-tuning pre-trained models, covering data preparation, tokenization strategies, loss functions, optimization algorithms, distributed training, and evaluation metrics. The curriculum explains how to structure training pipelines, handle different data formats, implement various fine-tuning approaches (full fine-tuning, LoRA, prompt tuning), and measure model performance. This includes both the mathematical foundations and practical implementation considerations for training at different scales.

Teaches systematic approaches to prompt design, few-shot learning, chain-of-thought prompting, and in-context learning strategies that improve LLM performance without model retraining. The curriculum covers how to structure prompts for different tasks, leverage examples effectively, use intermediate reasoning steps, and combine multiple prompting techniques. This includes both empirical best practices and theoretical understanding of why certain prompting strategies work better than others for different model sizes and capabilities.

Teaches how to build RAG systems that augment LLM generation with retrieved context from external knowledge sources, covering document indexing, retrieval mechanisms, ranking strategies, and integration with generation models. The curriculum explains how to structure knowledge bases, implement semantic search, handle retrieval failures, and optimize the retrieval-generation pipeline. This includes both the architectural patterns for RAG systems and practical considerations for production deployment with large document collections.

Teaches how to design autonomous agents that use LLMs for reasoning and decision-making, including planning algorithms, tool use and function calling, memory management, and multi-step task decomposition. The curriculum covers different agent architectures (ReAct, chain-of-thought, hierarchical planning), how to structure tool definitions for function calling, and strategies for handling agent failures and loops. This includes both the theoretical foundations of planning and practical implementation patterns for building reliable agents.

Teaches how to evaluate LLM performance across different dimensions including accuracy, fluency, factuality, safety, and efficiency, covering both automatic metrics and human evaluation methodologies. The curriculum explains how to select appropriate benchmarks, design evaluation protocols, interpret results, and understand the limitations of different metrics. This includes coverage of standard benchmarks (GLUE, SuperGLUE, MMLU, etc.), task-specific metrics, and emerging evaluation challenges for large models.

Teaches how to deploy LLMs in production environments with focus on inference optimization, latency reduction, and cost efficiency, covering quantization, distillation, batching strategies, caching, and hardware selection. The curriculum explains how to profile model performance, identify bottlenecks, implement optimization techniques, and measure trade-offs between quality and efficiency. This includes both software optimization techniques and hardware considerations for different deployment scenarios (cloud, edge, on-premise).

Teaches how to identify and mitigate risks in LLM systems including bias, toxicity, hallucination, and misuse, covering safety evaluation methodologies, alignment techniques, and responsible deployment practices. The curriculum covers red-teaming approaches, bias detection and mitigation, factuality verification, and ethical considerations in LLM development. This includes both technical safety measures and broader considerations for responsible AI deployment.

Teaches how multimodal LLMs process and generate content combining text, images, and other modalities, covering vision encoders, cross-modal alignment, and applications like image captioning and visual question answering. The curriculum explains how vision-language models integrate visual and textual information, the architectures used for multimodal fusion, and how to leverage multimodal capabilities in applications. This includes both understanding existing multimodal models and considerations for building or fine-tuning multimodal systems.