What can you build with reasoning-from-scratch?

Learn how GRPO reinforcement learning trains a model to reason by running it yourself. Compare chain-of-thought, self-consistency and self-refinement on MATH-500 problems. Distill a larger teacher model's reasoning into a smaller model using Ollama or OpenRouter outputs. Study a from-scratch Qwen3 implementation in PyTorch. Build a simple chat interface around a model you trained.

How do I get started with reasoning-from-scratch?

Clone the repository, then set up Python using the tips in Chapter 2. Open the ch02 notebook and work through the chapters in order. Commands from the README: git clone --depth 1 https://github.com/rasbt/reasoning-from-scratch.git.

rasbt /

reasoning-from-scratch

Code for a book that adds reasoning to a small Qwen3 model in PyTorch

5.4k stars
Jupyter Notebook
Updated 20 days ago
Some experience
Learning & resource lists

Open on GitHub (opens in a new tab) Website (opens in a new tab)

Reasoning models like DeepSeek R1 are easy to misread if you only read about them in theory. This repository is the official code for the book Build a Reasoning Model (From Scratch). You start with a pre-trained open base LLM, Qwen3, and add reasoning skills yourself in PyTorch.

The chapters go in order. You load and run the model, then build an evaluation setup on MATH-500. Next come inference-time methods such as chain-of-thought prompting, self-consistency, best-of-N and self-refinement. After that you train with reinforcement learning (GRPO), improve it, and distill reasoning from a larger teacher model.

The code is for learning, so the model is small and the main chapters are meant to run on consumer hardware. It stands alone, though the author's earlier book covers building a base LLM. The repo follows a printed book, so changes to the main chapter code are not accepted as contributions.

What it does

Chapter notebooks with solutions
Chapters 2 to 8 each have a main Jupyter notebook and a notebook with exercise solutions.
Inference-time scaling
Covers chain-of-thought prompting, self-consistency, best-of-N and self-refinement, with scripts that test them on MATH-500.
Reinforcement learning with GRPO
Shows how to train a reasoning model with GRPO, then improve it. Bonus scripts include DeepSeek-V3.2-, Olmo3- and GDPO-style training.
Distillation
Generates teacher outputs through Ollama or OpenRouter and trains a smaller model on them, with single-example and batched scripts.
Checkpoints and extras
Includes downloadable training checkpoints, Qwen3 source code, Hugging Face usage, evaluation methods and a chat interface.

What you could build

Learn how GRPO reinforcement learning trains a model to reason by running it yourself.
Compare chain-of-thought, self-consistency and self-refinement on MATH-500 problems.
Distill a larger teacher model's reasoning into a smaller model using Ollama or OpenRouter outputs.
Study a from-scratch Qwen3 implementation in PyTorch.
Build a simple chat interface around a model you trained.

Get started

Clone the repository, then set up Python using the tips in Chapter 2. Open the ch02 notebook and work through the chapters in order.

git clone --depth 1 https://github.com/rasbt/reasoning-from-scratch.git

Commands are copied from the project's README. Read it on GitHub for the full setup.

Who it's for

Suits students and engineers who know Python and basic deep learning and want to understand reasoning LLMs by coding them; a GPU helps for chapters 5 and 6.

Useful for

ML engineers and data scientistsImplement GRPO, inference-time scaling and distillation in PyTorch on a Qwen3 base model, following a book.
Students and beginnersWell-explained notebooks that build reasoning LLM techniques step by step, with code instead of just papers.
AI engineersCompare chain-of-thought, self-consistency and self-refinement on MATH-500 to understand how reasoning models work.