How the content is made

What ML Atlas is

ML Atlas is an encyclopedia of machine learning: 277 entries in Polish and 40 in English. Every entry has the same structure: what the concept is, the mechanism (why it works), an example on real data, practice, frequently asked questions and sources. Many of the Polish entries have interactive illustrations that compute on the same data in your browser.

Alongside it there is a machine learning course from zero: 33 lessons in 6 languages, each with a lab where you change the data or the model and see the result straight away.

Atlas is for people who want to understand how machine learning, neural networks and language models really work: from first steps in the course to entries on specific methods.

The lesson never lies

This is the rule every text starts from: every number on the site is either computed or taken from a cited source. This is how we keep to it:

• Examples in entries and lessons are computed on 9 public reference datasets (Iris, Titanic and the Palmer penguins among them). Many entries have a collapsible “How it was computed” box under the example with the settings used.
• When an entry describes what its illustration shows, a test checks that the numbers in the text are exactly the ones the illustration computes. Some tests also compare the illustrations’ results with scikit-learn.
• The code in the “In code (Python)” layer of every lesson was run and its output saved. A test checks that the lesson has the same code in every language and shows exactly that output.
• The lessons come with Jupyter notebooks that fetch the data from its public source themselves. A separate tool runs all of them the way a reader would and compares the result with the lesson. It needs the network, so it runs outside the automated tests.
• Numbers that cannot be computed on our data (such as results reported in papers) are given with their source.

Altogether that is over 1000 automated tests (as of 11 October 2026). They run on every change, together with a content validator and a site audit. The project rule: a change goes live only when all of them pass.

Sources

Every entry ends with a “Sources” section: the papers, textbooks and library documentation that the definitions, formulas and numbers come from. The content validator rejects an entry without it.

Data and copyright

Each of the 9 reference datasets is described on the Reference datasets page (in Polish): source, citation, licence and what we changed from the original. Most are under CC BY 4.0 or CC0. Where the authors give no licence (Titanic, Diabetes), we say so plainly.

The text is written from scratch, in our own words. We do not copy other people’s text, images or recordings. Short quotations, such as well-known statements of laws, are given with author and year. Charts and illustrations are graphics drawn by the site’s code from the data. The Inter font is under the SIL OFL 1.1.

Atlas content is licensed under CC BY-SA 4.0.

The role of AI

Atlas text and code are written with the help of language models (Claude by Anthropic). That is why we do not rely on what a model “remembers”: the numbers in examples are computed by code on the data, and every change passes tests and content rules before it is published, such as the required sections of an entry and numbers matching the illustrations.

News posts on the blog (in Polish) are not written from a model’s memory: each one lists its sources and carries a date.

The owner of Atlas sets the project’s rules and direction.