LLM Learning Curriculum

This repository contains a structured, hands-on curriculum for learning Large Language Models (LLMs) from the ground up. The materials are organized into a weekly, day-by-day plan that progresses from fundamental concepts to advanced topics. Each week's folder contains self-contained Jupyter notebooks with practical implementations of the concepts discussed.

---

Visual guide reference - https://bbycroft.net/llm

---

Week 1: Foundations

• Topics Covered: Text Normalization, Tokenization (BPE/WordPiece), Embeddings, and Positional Encodings.
• Outcome: Built a foundational understanding of how text is processed and prepared for a transformer model.
• Code: See the week1 directory for hands-on implementations.

Week 2: Transformer Architecture

• Topics Covered: Scaled Dot-Product Attention, Multi-Head Attention, Encoder-Decoder Architectures, and building a minimal transformer block from scratch.
• Outcome: Implemented the core components of the transformer architecture.
• Code: See the week2 directory, which includes a complete, minimal transformer implementation.

Week 3: Pretraining a Language Model

• Topics Covered: Language modeling datasets, perplexity, large-scale tokenization, sequence packing, and building a complete training loop with regularization (Dropout, Weight Decay), monitoring, and early stopping.
• Outcome: Developed a robust pipeline for pretraining a GPT-style model and implemented a capstone project inspired by nanoGPT.
• Code: See the week3 directory for a full pretraining implementation.

Week 4: Instruction Tuning & Alignment

• Topics: Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), and Direct Preference Optimization (DPO).
• Objective: Learn how to align a pretrained model to follow instructions and adhere to human preferences.

Week 5: Prompt Engineering & Retrieval-Augmented Generation (RAG)

• Topics: Advanced prompting techniques (Chain-of-Thought), function calling, and building RAG pipelines with vector stores and rerankers.
• Objective: Enhance model capabilities by integrating external knowledge and tools.

Week 6: Advanced Fine-Tuning Techniques

• Topics: Parameter-Efficient Fine-Tuning (PEFT) with LoRA and QLoRA, and knowledge distillation.
• Objective: Learn to efficiently adapt large models for specific tasks on consumer-grade hardware.

Week 7: Inference, Serving, and Optimization

• Topics: Model quantization, KV caching, and deploying models with high-throughput servers like vLLM and TGI.
• Objective: Understand how to optimize and serve LLMs for production environments.

Week 8: Agents, Tools, and Safety

• Topics: Building autonomous agents that can use tools, manage memory, and operate within safety guardrails.
• Objective: Develop a multi-step agent capable of solving complex tasks and completing the capstone project.