llm-course

Organizes LLM education into three progressive learning tracks (Fundamentals, Scientist, Engineer) with explicit entry points and dependency mapping, implemented as a single markdown hub that links to ~150+ external resources. Users navigate via a hierarchical section structure that maps learning paths to specific topics, with each topic following a consistent pattern of curated articles, videos, and tools. The architecture uses a documentation-first approach where the README.md acts as a central knowledge graph rather than containing executable code.

Aggregates 24 theoretical topics across three learning paths and embeds curated external references (articles, papers, videos, tools) directly within each topic section. Implementation uses a consistent topic section pattern where each topic links to 3-8 external resources selected for pedagogical value. The curation layer filters and organizes content from diverse sources (research papers, blog posts, YouTube, GitHub projects) into a single navigable structure without duplicating content.

Provides educational content on Retrieval Augmented Generation (RAG) and vector storage systems, covering vector databases (Pinecone, Weaviate, Milvus), embedding models, retrieval strategies, and advanced RAG techniques (re-ranking, query expansion, hybrid search). Content is organized as two dedicated sections within the LLM Engineer track and links to vector database documentation, embedding model resources, and RAG frameworks (LangChain, LlamaIndex). This capability enables practitioners to build knowledge-grounded LLM applications without fine-tuning.

Provides educational content on building LLM agents that can plan, reason, and use tools to accomplish complex tasks. Content covers agent architectures (ReAct, Chain-of-Thought), tool calling and function schemas, planning strategies, and agent frameworks (LangChain, AutoGPT, CrewAI). This capability is organized as a dedicated section within the LLM Engineer track and links to agent research papers, framework documentation, and implementation examples. Enables practitioners to build autonomous systems that go beyond simple prompt-response interactions.

Provides educational content on optimizing LLM inference for latency and throughput, covering techniques like batching, caching, quantization, and serving frameworks (vLLM, TensorRT-LLM, Ollama). Content is organized as a dedicated section within the LLM Engineer track and links to optimization papers, serving framework documentation, and performance benchmarks. This capability enables practitioners to deploy models efficiently and meet production latency/throughput requirements.

Provides educational content on deploying LLMs to production, covering containerization (Docker), orchestration (Kubernetes), cloud platforms (AWS, GCP, Azure), monitoring, and operational considerations. Content is organized as a dedicated section within the LLM Engineer track and links to deployment frameworks, cloud documentation, and best practices. This capability enables practitioners to move models from development to production with proper infrastructure, monitoring, and reliability patterns.

Provides educational content on securing LLM applications and addressing safety concerns, covering prompt injection attacks, data privacy, model poisoning, adversarial robustness, and compliance considerations. Content is organized as a dedicated section within the LLM Engineer track and links to security research, safety frameworks, and best practices. This capability enables practitioners to build LLM applications with appropriate security and safety guardrails.

Provides educational content on evaluating LLM quality and performance, covering automatic metrics (BLEU, ROUGE, BERTScore), human evaluation, benchmarks (MMLU, HellaSwag, TruthfulQA), and evaluation frameworks. Content is organized as a dedicated section within the LLM Scientist track and links to evaluation papers, benchmark datasets, and evaluation tools. This capability enables practitioners to measure model quality and compare different models or training approaches.

Provides curated content on emerging LLM techniques and research trends, covering recent advances in model architecture, training methods, and applications. Content is organized as a dedicated section within the LLM Scientist track and links to recent research papers, blog posts, and tools implementing new techniques. This capability enables practitioners to stay current with rapidly evolving LLM field and understand cutting-edge approaches.

Provides 23 executable Jupyter notebooks hosted on Google Colab that implement theoretical concepts from the course, organized into four categories: Automated Tools (8), Fine-tuning (6), Quantization (4), and Advanced Techniques (5). Notebooks are embedded as links within relevant course sections, creating a tight coupling between theory and practice. Each notebook implements specific techniques (e.g., LoRA fine-tuning, GGUF quantization, model merging) with runnable code that requires only a Google account and GPU access.

Provides an optional foundational learning path covering Mathematics for Machine Learning, Python for Machine Learning, Neural Networks, and Natural Language Processing. This track is marked as optional (not required for advanced learners) and spans lines 74-157 of the README, serving as a prerequisite for both Scientist and Engineer tracks. Implementation uses a modular topic structure where each fundamental topic links to external resources (textbooks, courses, tutorials) rather than providing original content.

Provides a core learning path (8 topics, lines 159-304) focused on building and training LLMs from scratch, covering LLM Architecture, Pre-Training Models, Post-Training Datasets, Supervised Fine-Tuning, Preference Alignment, Evaluation, Quantization, and New Trends. This track is designed for researchers and practitioners wanting to understand model internals and training pipelines. Implementation uses the same topic-curation pattern as Fundamentals but with deeper technical content (papers, research blogs, training frameworks).

Provides a core learning path (8 topics, lines 305-440) focused on deploying and operating LLMs in production, covering Running LLMs, Vector Storage, Retrieval Augmented Generation (RAG), Advanced RAG, Agents, Inference Optimization, Deployment, and Security. This track is designed for engineers building LLM applications and systems. Implementation uses the same topic-curation pattern but emphasizes tools, frameworks, and operational concerns over research papers.

Provides comprehensive educational material on transformer architecture fundamentals, covering decoder-only architectures (GPT, Llama, Mistral), tokenization methods, attention mechanism variants, and text generation strategies. Content is organized as a dedicated section within the LLM Scientist track and links to foundational papers (Attention is All You Need), implementation guides, and visual explanations. This capability serves as the architectural foundation for understanding all downstream topics (pre-training, fine-tuning, quantization).

Provides educational content on pre-training LLMs from scratch and curating post-training datasets, covering model initialization, training objectives (next-token prediction, masked language modeling), dataset composition, and scaling laws. Content is organized as two dedicated sections within the LLM Scientist track and links to research papers (Chinchilla, Scaling Laws), dataset resources (Common Crawl, Wikipedia), and training frameworks (Hugging Face Transformers, Megatron). This capability bridges architecture understanding with practical training considerations.

Provides educational content and 6 executable notebooks on supervised fine-tuning (SFT) and preference alignment techniques (RLHF, DPO, IPO). Content covers fine-tuning methodologies, dataset preparation, and alignment algorithms, with notebooks implementing LoRA fine-tuning, full fine-tuning, and preference alignment on Colab. This capability enables practitioners to adapt pre-trained models to specific tasks and align outputs with human preferences without requiring massive compute.

Provides educational content and 4 executable notebooks on quantization techniques for reducing model size and inference latency, covering post-training quantization (PTQ), quantization-aware training (QAT), and specific formats (GGUF, GPTQ, AWQ). Content links to research papers on quantization methods and includes notebooks implementing quantization pipelines on Colab. This capability enables deployment of large models on resource-constrained hardware without significant quality loss.