Compendium of free ML reading resources: a 197-book link list and what it costs you

A reading list, not a library: what the repository actually is

The repository is a single index. The README opens with a key that defines two markers, an orange book for a PDF file and an earth globe for an HTML book, then a table of contents listing 18 topics with a count in brackets next to each. The counts run from 1 for anomaly detection to 51 for machine learning, and the README states the total number of books as 197. Below the table of contents, each topic heading is followed by a bulleted list of entries in the form of a marker, a quoted title, a link, and an author line. There is no code in the repository that the supplied material describes, no build step, no data file, and no retrieval mechanism. The primary language is listed as unknown and no licence is given. That is the whole artefact: a Markdown file whose usefulness is entirely a function of whether the titles and links are good. For an engineer deciding whether to adopt it, the honest framing is that this is not a tool you install. It is a document you read, and possibly a document you send to someone else.

The problem it solves: where to start when every ML syllabus is a paid course

The stated scope is free material. The README carries a note that all books listed have been made freely available by their respective authors or publishers, and that arXiv papers are inherently free. That constraint is the point. A person who wants to learn the mathematics behind a model, or read a full treatment of Bayesian data analysis, faces a market where the default answer is a paid textbook or a course subscription. This list answers a narrower question: given that free material exists, which specific items are worth the time. The topic counts show where the compiler thinks the weight lies. Machine learning gets 51 entries, statistics and probability 16, deep learning 17, R-related material 18, causal inference 14, and NLP and large language models 13. The long tail is thin: anomaly detection has 1, recommender systems 2, feature engineering 2, computer vision 3. That distribution is a signal about the compiler's interests, and it is worth reading as one. If your work is in computer vision, this list will not carry you far.

How the index is structured and how you consume it

The mechanism is a two-level hierarchy with no metadata beyond format. Level one is the topic heading, and each heading is an anchor target that the table of contents links to, for example the entry for EDA, visualization, and data cleaning points at the anchor #eda-visualization-and-data-cleaning. Level two is the bullet. A typical bullet reads as the HTML marker, a quoted title, a link to the host, and the author names. The mathematics section, for instance, lists Mathematics for Machine Learning by Deisenroth, Faisal, and Ong as a PDF link to mml-book.github.io, and Linear Algebra Done Right by Axler as a PDF link to a Springer content URL. The statistics section lists Bayesian Data Analysis by Gelman and co-authors as a PDF hosted on a Columbia University page. The data flow is trivial: you scroll or jump to a topic, scan titles, click a link, and land on a third-party host. There is no search, no tagging by difficulty, no prerequisite graph, and no indication of length. The format markers do tell you something useful before you click: a PDF means a download, an HTML book means you read in the browser. That distinction matters on a locked-down work machine, and it is the only piece of per-entry metadata the list provides.

Getting it: clone, read, and what the README gives you

There is nothing to install. The repository has a default branch named main and no homepage. The commands are the ordinary ones for a Markdown repository. To get a local copy, run git clone https://github.com/Carl-McBride-Ellis/Compendium-of-free-ML-reading-resources and then read README.md in the editor of your choice. If you only want the raw file, the raw URL form on the main branch returns the same content. There are no config keys, no environment variables, and no entry points, because there is no executable component in the material supplied. The absence of releases is consistent with that: the repository has no releases retrieved, and the last push recorded is 2026-09-04. Practically, the only workflow this supports is reading. If you want the list in another form, you would have to convert the Markdown yourself, and because entries are free-form bullets rather than a structured table, any automated parsing has to tolerate inconsistent ordering of marker, title, link, and author. That is a real cost if you were hoping to feed this into a reading tracker.

Link rot is the failure mode, and the list cannot detect it

Every entry depends on a host the compiler does not control. The mathematics and statistics sections are the clearest illustration: they point at university course pages, personal academic sites, and a Google Drive file for at least two entries in the linear algebra list. A Google Drive link is the most fragile kind of citation in a public index, because access can change with a sharing setting that has nothing to do with the book. University pages are the second most fragile, since faculty move and course directories get reorganised. A static Markdown file has no way to notice. There is no CI job described in the material, no link checker, and no release cadence that would surface breakage. The consequence for a reader is that the list degrades silently over time, and the only repair path is a pull request against the single README. A second limitation is coverage bias. The counts are not a survey of the field; they reflect what one compiler found and chose. Anomaly detection at 1 entry and feature engineering at 2 entries are not claims that little free material exists, they are claims about what made it into this file. Treat the counts as an inventory, not a ranking.

Compared with Awesome Machine Learning: curation versus cataloguing

The obvious alternative in the same genre is the Awesome Machine Learning list, which is a catalogue of software libraries, frameworks, and tools organised by language and task, with the standard Awesome list conventions and contribution guidelines. The difference in approach is what gets indexed. Awesome Machine Learning answers the question of what to build with; this compendium answers the question of what to read. That distinction determines who each list serves. If you are choosing a gradient boosting library or looking for a Python package to do text classification, the Awesome list is the right shape and this compendium has nothing for you, since its entries are books and papers rather than repositories. If you are trying to understand why a random forest behaves the way it does, or you want a full treatment of conformal prediction beyond a blog post, this list is the better starting point because it points at long-form explanations. The trade-off is maintenance. Awesome lists typically carry contribution rules and a defined entry format, which makes automated checking feasible. This compendium carries no such rules in the supplied material, so entries can drift in shape as well as in link health.

Maintenance, licence, and what you are actually taking on

The maintenance burden falls on whoever wants the list to stay accurate. The last push is recorded as 2026-09-04, and there are no releases, so there is no version to pin and no changelog to read. Upgrading means pulling the latest README and diffing it against your copy, which is a manual step. If you embed the list in internal documentation, you inherit the link rot without inheriting any notification. On licensing, the material does not state a licence for the repository, which is a gap worth noting plainly: without a stated licence, the terms under which you may redistribute the list are unclear, and that is a question for whoever owns the policy, not something to assume. The books themselves are a separate matter entirely. The README asserts that the listed books were made freely available by their authors or publishers, but free to read is not the same as free to redistribute or to use commercially, and the list does not record the licence of any individual book. If you plan to use a title as course material, check that title's own terms rather than relying on its presence here.

Editorial conclusion

Adopt this if you want a starting curriculum for self-study or a reading list to hand to a junior engineer, and you are willing to accept links you cannot control. Do not adopt it if you need a versioned, reproducible dependency, per-book reviews, or anything with a maintenance guarantee. Before relying on it, open a few of the PDF links in the mathematics and statistics sections, since those are the ones most likely to point at university or personal hosts that move; if a link is dead, the fix is a pull request against the single README.md, not a fork.

Official sources

1. Carl-McBride-Ellis/Compendium-of-free-ML-reading-resources on GitHub
2. Issues
3. README