mlabonne/llm-course

LLM Course is a free, self-paced course that teaches how to work with Large Language Models, the kind of AI behind ChatGPT and similar tools. It is not a piece of software you install but a learning resource. The repository is mostly markdown roadmaps, links to articles, and Google Colab notebooks you can open in your browser and run without setting anything up locally. The course is split into three tracks. The first, called LLM Fundamentals, is optional and covers the background (the mathematics, the Python, and the basics of neural networks) that the later tracks assume. The second, The LLM Scientist, focuses on building the best possible language models using current techniques. The third, The LLM Engineer, focuses on the practical side of taking an existing model and turning it into an application you can deploy. Each track is laid out as a roadmap with notebooks attached, so the theory comes with runnable examples. Alongside the roadmaps, the repository hosts a library of notebooks for specific tasks. There are notebooks for fine-tuning models like Llama and Mistral using methods such as QLoRA, ORPO, and DPO, for quantising models in formats like GGUF, GPTQ, EXL2, AWQ, and HQQ so they run on smaller hardware, for merging models with MergeKit and visualising the resulting model family tree, for evaluating models automatically with LLM AutoEval on RunPod, and for spinning up a Gradio chat interface for free with ZeroGPU. There are also helper notebooks for dataset deduplication and for ablation-style model edits. Someone would use this course to move from curious about LLMs to actually training, modifying, and shipping one, while reading short blog-post-style articles alongside each notebook. The author, Maxime Labonne, has also co-written a paid book called the LLM Engineer's Handbook that covers a full LLM application from design to deployment, but the course itself is stated to remain free. A community-generated DeepWiki version is linked for readers who want a more browsable walkthrough.