Skip to main content
You can read Python and run a script. You have never trained a model, and the words in the quickstart (rollout, reward, held-out, pass@1) are a jump. These seven lessons close that jump one word at a time. Each lesson says the idea in plain words first, then names the term the field uses for it, then runs it. Every snippet runs on a laptop with no key and no network, and the output shown is what it printed. At the end you will be able to read the quickstart and know what every line is for.
1

A model learns from examples, not from instructions

What training changes that a prompt cannot. Pretraining, post-training, SFT and RL, in one page. Read it.
2

One saved conversation is a rollout

The row: the ask, the tool calls, the reply. Where rows come from when you have no traffic. Read it.
3

A reward is a score you can defend

A program that checks the answer, or a model that reads it. How to check the checker. Read it.
4

A pass rate without an interval is a guess

pass@1, the interval, and how many tasks a real result needs. Read it.
5

The test the model never sees is the only score that counts

The held-out set, and the check that no training row is a copy of a test row. Read it.
6

Train on what the model gets right sometimes

Which rows to keep for SFT, which for RL, and why a task the model always passes teaches nothing. Read it.
7

Training is done when the held-out score moved

Export, train on your own compute, and prove the change with a paired before and after. Read it.

​
What you need

Python 3.10 to 3.13. No account, no key. The lessons use a stand-in agent that makes mistakes on purpose and records them, so there is always an answer key to check against.

​
The whole course in one paragraph

A language model has read the internet and never worked at your company. A prompt is a note on its monitor. Training changes what it knows how to do. To train it you need rows (one saved conversation each), a score per row (a program if one can check, a model if not, and you check the model), and a test set the model never sees. You keep the rows that teach something, train on your own compute, and the result counts only when the score on the test set moves by more than the noise. Then the next day’s traffic is the next training set.
Last modified on September 19, 2026