mlabonne/llm-course is the LLM curriculum the labs won't write down

Maxime Labonne’s llm-course repository does not teach you to train a frontier model. It teaches you to take one apart, reweight it, compress it, merge it with a neighbor, and run it on a laptop. The course splits into three parts: an optional LLM Fundamentals section on math, Python, and neural networks; an LLM Scientist track on building models; and an LLM Engineer track on deploying applications. The Scientist track is where the interesting material lives, and it is almost entirely operational. Fine-tune Llama 3.1 with Unsloth. Quantize to GGUF with llama.cpp. Merge models with MergeKit, no GPU required. Strip refusals with abliteration.

That last item is the tell. A course that ships a notebook called “Uncensor any LLM with abliteration” is not a course in the academic sense. It is a toolchain with a syllabus attached.

The notebooks are the product

The repository’s real content is a table of Colab notebooks, each paired with a written article. Fine-tuning Llama 3 with ORPO, a single-stage method that folds supervised fine-tuning and preference alignment together. Fine-tuning Mistral-7B with DPO and QLoRA, the latter explicitly sized for a free-tier Colab instance. Quantization in four flavors: 8-bit, GPTQ at 4-bit, GGUF via llama.cpp, and EXL2 through ExLlamaV2. Model merging with MergeKit, including a recipe for building “frankenMoE” mixtures by combining multiple experts into one model.

Then the automation layer. LLM AutoEval runs evaluations on RunPod. LazyMergekit merges models in one click. LazyAxolotl fine-tunes in the cloud in one click. AutoQuant converts to GGUF, GPTQ, EXL2, AWQ, and HQQ in one click. AutoDedup strips duplicate rows from training sets using the Rensa library. ZeroSpace spins up a Gradio chat interface on free ZeroGPU.

Read the list again and notice what it describes: a pipeline. Data cleaning, fine-tuning, alignment, merging, quantization, evaluation, deployment. Every stage has a one-click wrapper. The course is not teaching concepts in the abstract. It is handing over a working assembly line for derivative models, and it charges nothing for it.

Why this matters more than it looks

The open-weight ecosystem runs on exactly this kind of derivative work. Most of the models on Hugging Face are not pretrained from scratch. They are fine-tunes, merges, and quantizations of a small number of base checkpoints, mostly Meta’s Llama family and Mistral’s releases. The skills the llm-course teaches are the skills that produce that long tail.

That has a downstream effect on compute economics. A 4-bit GPTQ or GGUF quantization lets a 7B or 8B model run on consumer hardware, which is the difference between an inference bill and no inference bill. ExLlamaV2 exists because people wanted faster local generation, not because a lab needed it. The course treats quantization as a first-class engineering discipline rather than an afterthought, and that framing is correct. For a large share of real deployments, the quantization step matters more to the unit economics than the base model choice does.

It also has a policy edge. Abliteration, the technique of removing refusal behavior through fine-tuning without full retraining, appears here as a tutorial. That is a genuinely contested capability. It sits at the center of the argument over whether open-weight releases can be safety-tuned at all, since a refusal direction baked into weights can apparently be subtracted back out. The course does not editorialize. It just shows the notebook. Anyone who wants to argue that open weights are safe by default has to reckon with the fact that the removal recipe is public, documented, and one click away.

The monetization is honest, which is rare

Labonne does not hide the business model. The course stays free, and he co-wrote the LLM Engineer’s Handbook, a paid book that walks an end-to-end LLM application from design to deployment. The repository says so plainly at the top. There is also a DeepWiki build for a more comprehensive version.

Contrast that with the paid “AI engineer” bootcamps that repackage the same public material behind a paywall and a cohort schedule. The llm-course undercuts them by existing. Its 3Blue1Brown and Khan Academy links in the Fundamentals section are the same free resources everyone else cites. The difference is that Labonne assembled them into a sequence and then attached runnable code to each stage.

A course that ships a notebook for stripping refusals out of a model is not a course in the academic sense. It is a toolchain with a syllabus attached.

What the course does not cover

Worth naming the gaps, because the framing implies more than it delivers. The Fundamentals section is explicitly optional and pitched at people who need a refresher, not at people learning linear algebra for the first time. The Scientist track teaches you to fine-tune and merge, not to pretrain. There is nothing here on distributed training across thousands of accelerators, nothing on data mixture design at pretraining scale, nothing on the infrastructure that gets a cluster to stay up. Those are the skills that separate a derivative-model practitioner from a frontier-lab researcher, and they are not in the notebooks.

The course is also, by design, tied to specific tools and versions. Llama 2 and Llama 3 appear as fine-tuning targets. Mistral-7B appears twice. Those checkpoints age. The methods (QLoRA, DPO, ORPO, GPTQ) have longer shelf lives than the models they are demonstrated on, and the repository will need steady maintenance to stay current. Whether it gets that maintenance is an open question. Labonne’s attention is now split across a book, a blog, and the notebooks themselves.

What to watch

The interesting signal is not the star count. It is whether the one-click wrappers keep pace with the methods. If AutoQuant adds new quantization formats as they land, and if the fine-tuning notebooks migrate to whatever replaces LoRA-style adapters, the course stays a live reference. If they stall on Llama 3-era recipes, it becomes a historical document.

For builders, the practical takeaway is narrower. If you are shipping a model that is a fine-tune, a merge, or a quantized build of someone else’s checkpoint, the pipeline in this repository is close to the current default. Read the notebook before you read the paper. The notebook is what actually runs.