tensorflow/docs: the source files behind tensorflow.org

What tensorflow/docs actually contains

The repository is the source of the guide and tutorial pages published on tensorflow.org. The README states this directly: these are the source files for the guide and tutorials on tensorflow.org. Most of the content is Jupyter Notebook, which is also the primary language listed for the repository, and the top-level layout is small: site/, tools/, setup.py, CONTRIBUTING.md, CODEOWNERS, MANIFEST.in, AUTHORS and the licence file.

The audience is narrow. This is not a library you import to train a model. It is for people who write, review or translate TensorFlow's own documentation, and for anyone who needs the tensorflow-docs package, which setup.py describes as a package for generating python api-reference docs. If you want runnable TensorFlow code, the notebooks here are documentation source, not a distribution channel for the framework.

How the pipeline splits between site/ and tools/

The split is the clearest architectural signal in the repository. site/ holds the published content, including the language directories used for localised docs. tools/ holds the Python package, and setup.py derives its version number from the timestamp of the last commit that touched tools/. That is a CalVer scheme, and the file cites calver.org in a comment. The consequence is that the package version is not a curated release number but a function of git history, so a shallow clone or an export without git metadata breaks the version computation.

The dependencies reflect what the generator does, not what TensorFlow does. REQUIRED_PKGS lists astor, absl-py, jinja2, nbformat, protobuf>=3.12 and pyyaml. A separate VIS_REQUIRE group adds numpy, PILLOW and webp for visual output. There is no tensorflow dependency in either list, which is the point: this toolchain reads notebooks and docstrings and renders pages, and needs no accelerator to do it.

Translations are out of scope here. The README says community translations live in the tensorflow/docs-l10n repository and are contributed, reviewed and maintained by the community as best-effort. If you are chasing a localised page, the file may not be in this repository at all.

Installing tensorflow-docs and confirming the checkout

setup.py sets the distribution name to tensorflow-docs and requires Python 3.9 or newer, so check your interpreter first. The README does not give a pip command, but the packaging metadata is standard setuptools, so an editable install from a clone is the normal route.

python -m pip install -e .

Run that from the directory containing setup.py. The version string is computed by shelling out to git log -1 --format=%ct tools, so the command needs a real git checkout rather than a downloaded archive. With truncated history or no git binary, the install fails while generating metadata, not later at import time.

After installation the package is used to generate API reference pages from source docstrings. The setup.py docstring identifies it as a package for generating python api-reference docs, and the jinja2 and nbformat dependencies show that templates and notebook parsing are the two moving parts. The README does not document a single canonical build command, so treat the notebooks under site/ as the editable artefacts and the tools/ package as the renderer you invoke around them.

import tensorflow_docs
print(tensorflow_docs.__file__)

That prints the path the import resolved to. If it points under site-packages rather than into your clone, the editable install did not take effect.

Where this repository will not help you

The most common mismatch is expecting documentation source to behave like a product. The README documents no CLI, no published build target and no rollback procedure, and it does not explain how to preview a page locally. That work is described in CONTRIBUTING.md and in the external TensorFlow docs contributor guide, both of which the README points to instead of restating.

Issue routing is the second boundary. The README sends docs issues to the issue tracker in the tensorflow/tensorflow repository using a documentation issue template, not to this repository's own tracker. A typo report filed here lands in the wrong queue.

The third is scope drift. This is not the place to report a bug in TensorFlow itself, and it is not where framework release notes live. It contains prose and notebooks about TensorFlow plus the tooling that renders them. Anyone installing tensorflow-docs hoping for a working TensorFlow runtime will find no such dependency, by design.

tensorflow/docs compared with a hosted docs platform

The obvious alternative for a documentation set of this shape is a hosted service that builds from a repository on every push, usually driven by a config file and a theme. Those systems trade control for convenience: you get a preview URL and a search index without writing a renderer, but you accept their template model and their handling of notebooks.

The difference here is that tensorflow/docs ships its own renderer as an installable package. The jinja2 and nbformat dependencies mean the page templates and the notebook parsing live in the repository and are versioned alongside the content. That is more setup work, which is why the README defers to contributing guides rather than offering a one-line preview command. In exchange, the published output is reproducible from the same commit as the source, and API reference pages are generated from docstrings rather than hand-written. For a documentation set this large, with API reference and translated variants, a hosted platform would move much of that generation logic outside the repository.

Maintenance, licensing and the cost of tracking upgrades

The last push to the default branch was on 2026-07-09, so the repository is not dormant. The only listed release is 2023.5.24.56664, marked as a first release and dated 2023-05-25. That gap follows from the versioning scheme, where setup.py computes the version from the last commit timestamp of tools/ rather than from tagged releases. Anyone pinning tensorflow-docs by version number is pinning a date, and every commit under tools/ produces a new one.

The upgrade cost is low but non-obvious. No changelog appears in the repository layout, and the version number tells you when tools/ last changed, not what changed. If you depend on the generator, follow the tools/ directory rather than the version string.

The licence is Apache-2.0, declared in the LICENSE file and reflected in the Apache header at the top of setup.py. It is permissive, permits commercial use and modification under its terms, and carries notice and attribution obligations plus a patent grant. That is a description of the licence text, not legal advice; read LICENSE and consult counsel if the stakes are high.

Editorial conclusion

Adopt tensorflow/docs if you write, review or translate the official TensorFlow guides and tutorials, or if you need the tensorflow-docs package that generates Python API reference docs from docstrings. Do not adopt it expecting runtime machine learning code, and do not file docs issues here: the README routes them to the tensorflow/tensorflow tracker. Verify first that your Python is 3.9 or newer, since setup.py sets python_requires='>=3.9', and confirm whether the page you want to edit lives under site/ or in the separate tensorflow/docs-l10n repository.

Frequently asked questions

Official sources

1. License: Apache-2.0
2. Project website
3. README
4. Releases
5. tensorflow/docs on GitHub