Construire un LLM de zéro
A free, French-language series of 22 notebook-based chapters that build a language model entirely by hand, from basic vector arithmetic through attention, the transformer architecture, and a reasoning agent.
Each of the 22 chapters pairs a notebook containing complete, runnable lesson code with a page on the project's own website carrying the explanatory text and interactive visualizations, such as watching attention concentrate or loss decrease as a slider is moved. A chapter's notebook ends with a graded set of exercises marked with TODO comments and validated by assert cells, and a matching solution notebook in a separate folder gives the worked answer once a learner is stuck or wants to check their work.
Topics move from the Python needed for the material, through autograd, randomness, neural networks, tokenization, embeddings, attention and the transformer architecture, up to training practices and, by the final chapters, agents that reason. Notebooks can be run in Google Colab with no local setup, or locally with a Python virtual environment and Jupyter Lab.
It is written as Jupyter Notebooks and released under the Apache-2.0 license, and the notebooks themselves are stated to remain free.