CookLLM

CookLLM is an educational platform for mastering large language models (LLMs) by building every core component from scratch, with interactive documentation and a hands-on code repository.

What is CookLLM?

CookLLM is a self-paced, project-based course that teaches the full stack of LLM development—from tokenization and Transformer architecture to distributed training and post-training alignment. It takes learners through two parallel tracks: a principle deep-dive series covering six major topics (foundations, systems, scaling, data, evaluation, and alignment) and a hands-on training track where you build and train a small model (BentoLM 29M–134M parameters) end-to-end. The platform is a website (cookllm.com) created by an independent team and includes a private GitHub repository for all code.

Key Features (or What makes CookLLM stand out?)

• From-scratch implementations — Every module (RustBPE tokenizer, Transformer, RoPE, Muon optimizer, RLHF loop) is built without relying on high-level framework wrappers, ensuring no black boxes.
• Interactive documentation — Formulas, diagrams, and executable code are combined in a single page; you can read and run code side by side.
• Two-codebase structure — The recipes repository contains principle-focused code snippets, while bento provides a complete training pipeline for a small LLM (data download → tokenizer training → pretraining → SFT → RLHF/PPO).
• Coverage of production concerns — Topics include GPU kernel optimization, distributed parallelism, scaling laws, data cleaning/dedup, and inference acceleration.
• Lifetime updates — A one-time purchase grants permanent access to all current and future chapters and the private GitHub repositories.
• Community support — An active Discord server (discord.gg/dKxBk7f9KB) for discussion and help.

Who is it for?

• AI engineers who want to move beyond API-calling and debug training issues at the kernel or architecture level.
• Students and researchers who need to translate paper formulas into working code (e.g., implementing RoPE or GRPO from first principles).
• Self-taught practitioners lacking formal ML education but wanting a structured, code-first understanding of LLM internals.
• Interview candidates seeking to answer deep implementation questions (e.g., “how does backprop through attention work?”) with confidence.

What can you do with CookLLM? (or Use cases)

• Build a custom tokenizer — Train a BPE tokenizer using RustBPE or tiktoken, then use it in your own model.
• Pretrain a small LM from scratch — Run the full BentoLM pipeline (29M–134M parameters) on your own data, including mid-training context extension and RoPE scaling.
• Implement alignment techniques — Walk through supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and direct preference optimization (DPO) with code.
• Understand scaling choices — Use the Scaling Laws module to compute optimal model size and data budget given a compute budget, then compare with actual training runs.

Pricing

CookLLM is a paid product with a single Lifetime tier offered at a discounted price of $39 (originally $139). This one-time fee unlocks all current and future course content, the private GitHub repositories, and early access to beta features. A free trial is available to get started before purchasing.

FAQ

Do I need a GPU to follow the course?

Not necessarily. The earlier principle modules run on CPU. For the hands-on training track, a small GPU (e.g., NVIDIA T4 or better) is recommended but you can also train the smallest model (29M parameters) on a modern laptop CPU with patience.

How does the “principle deep-dive” differ from the “hands-on training” track?

The deep-dive (recipes) covers each topic with focused, isolated code snippets and interactive visualizations; the hands-on track (bento) integrates everything into a single repository where you train and align a real model from start to finish.

Will content be added after purchase?

Yes, the platform promises lifetime free updates, including new chapters on advanced topics (e.g., Agent systems, MLLM) as they are developed.

Is the course available in English?

The website is primarily in Chinese but offers an English toggle. The code and documentation are code-centric, so language barriers are minimal for technical content.