Free course shows how to train reasoning models from scratch

Now it is possible for almost anyone to build their own reasoning model.

Sebastian Raschka has released a completely free course and opened the source code for all materials.

Using a small Qwen3-0.6B model as an example, he shows how reasoning models are actually trained and what stages make up the process.

The course covers how to:

• teach a model to search for different ways to solve the same problem;
• automatically verify the correctness of generated answers;
• improve model behavior through reward-based training and GRPO;
• observe how a model sometimes finds ways to bypass the reward system itself;
• train a more compact model on reasoning data from DeepSeek-R1.

All code can be run independently, allowing you to watch the training process and gradually notice how the model’s behavior changes.

You do not have to start from scratch: the course uses an already available small model as the starting point, rather than a foundation model that would need to be trained from zero.

If you have long wanted to understand how models like DeepSeek-R1 work, this course offers a rare opportunity to go through the entire process hands-on.

GitHub: reasoning-from-scratch