tutorials-kr: the PyTorch tutorials in Korean, archived per version
A repository for Korean translations of the official PyTorch tutorials.
At a glance
- What is it?
- A translation fork of the official PyTorch tutorials that builds as a Sphinx site at tutorials.pytorch.kr, pins the upstream commit it translates, and keeps every released version readable by tagging the first and last release of each. It asks for three kinds of help, and says which one it is short of.
- Who is it for?
- tutorials-kr is a good answer to a narrow and real problem, which is that translated framework documentation rots the moment upstream moves. Pinning the upstream commit and tagging the first and last release of each version means a reader on an older PyTorch can still read the Korean text that matched it.
- Can I use it commercially?
- Yes. BSD-3-Clause is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 110 days ago.
- What is it written in?
- Mainly Jupyter Notebook, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- PyTorchKR/tutorials-kr374 ★ · -2
| Repository | 2026-09-13 | 2026-11-15 |
|---|---|---|
| PyTorchKR/tutorials-kr | 376 | 374 |
-2 Stars in the last 63 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
A fork that pins the upstream commit it translates
The repository does not track a branch of upstream tutorials, it tracks a commit. The English footer states it plainly: this is a project to translate a specific upstream commit of pytorch/tutorials into Korean, and anyone who wants the latest version should go to the official repository.
That is the whole design decision, and it is the correct one for documentation. A translation is only meaningful relative to the text it was made from, so a moving target produces a site that is subtly wrong in ways nobody can audit.
The target at the time of writing is PyTorch v2.8, based on commit c4d9d93.
The repository is a Sphinx documentation project rather than a Python package. The top level carries conf.py, index.rst and a family of other .rst entry points such as deep-dive.rst, distributed.rst, extension.rst, ecosystem.rst, intro.rst and recipes_index.rst, plus the content directories beginner_source/, intermediate_source/, advanced_source/, recipes_source/ and unstable_source/, a _static/ directory, and docs/.
Two files explain the relationship to upstream without needing prose. A .gitmodules entry is how upstream content is brought in, and .migration_state.json is the kind of file a project keeps when it needs to track which upstream changes it has already absorbed.
CNAME at the top level is the custom domain declaration, which is what makes the build land on tutorials.pytorch.kr rather than a project page URL.
base and latest tags turn a fork into a version archive
The version scheme is the part worth understanding, and it is explained in the readme with examples.
A tag named version-base, for example 1.9-base, is the release made when that version of PyTorch started. A tag named version-latest, for example 1.9-latest, is the last release of that version. So each supported PyTorch version has two entries: where its translation began and where it ended.
From v1.0 onwards the translations are not managed in a separate repository. You find them in this repository's release menu, download the documents for the release you want, build them locally, and read the documentation as it stood for that PyTorch version. The build steps point at a section of the contributing guide rather than repeating them.
Older material is handled the other way, in dedicated repositories that are no longer translated: 0.4.1 and 0.3.1 each have both a documentation site and a GitHub source repository linked in a table. So there are three eras with three different arrangements, and the readme says which one applies to what.
The most recent tags follow the same pattern with different numbers: 2.8.0-base, 2.3.1-latest and 2.3.1-base, the two from the 2.3.1 era sharing a name.
That last detail is worth noticing, because a reader looking for 2.3.1 documentation has to pick between the first and last release of that version, and the readme does not say they are interchangeable.
Three ways to help, and the third one is the bottleneck
The contributing section lists three paths and marks one of them with a star.
The first is fixing a typo or improving an existing translation: you find a mistake on the Korean tutorials site and correct it in the Korean tutorials repository. This is the smallest unit of work and the easiest to start.
The second is translating a tutorial that has not been translated yet on the site.
The third is reviewing a document that already has a pull request against this repository, checking whether the translation is appropriate. It carries a star and a note that the participation of many people is warmly awaited, which in practice means this is the role the project is short of.
That asymmetry is worth taking at face value. A translation project accumulates pull requests faster than it accumulates reviewers, and an unreviewed translation is worse than an untranslated page because a reader cannot tell which is which.
The detailed method for all three is in the contributing guide, and the readme defers to it rather than duplicating the steps.
Translations reach the site irregularly
One sentence sets expectations for anyone reading the site: the results of the translation are visible at tutorials.pytorch.kr, and they are folded in irregularly as the translation proceeds.
That is an honest warning and it changes how you should use the site. The repository is where the work is current; the published site is a snapshot that moves when someone publishes it. Anyone comparing the Korean text against the English original and finding a difference may be looking at a site that has not caught up rather than a translation error.
Issues about translations and changes for the current version are tracked in a single issue, number 972, rather than as separate per-tutorial issues. That is a reasonable pattern for a translation fork: one place to see what is in flight for the version being worked on.
The site address is served over plain HTTP in the links given throughout the readme, which is worth noting for anyone publishing derived content from it.
Together with the pinned upstream commit and the base and latest tags, the picture is of a project that is deliberate about what it can promise: a specific upstream state, a tracking issue, and an irregular publish cadence.
Every unknown make target is routed to Sphinx
The Makefile is a Sphinx makefile in the upstream style, with a catch-all that is worth reading.
Variables first: SPHINXOPTS is overridable, SPHINXBUILD is sphinx-build, SPHINXPROJ is PyTorchTutorials, SOURCEDIR is the repository root, BUILDDIR is _build and DATADIR is _data. The locale is forced with an exported LC_ALL=C, which is standard for documentation builds and prevents translated error messages from changing behaviour.
Then the routing. A help target calls sphinx-build with the help mode, and a catch-all pattern rule passes any unrecognised target straight through to sphinx-build's make mode with that target name. That is why html, linkcheck and similar Sphinx builders work without being listed.
The consequence is that a typo becomes a Sphinx error rather than a make error. Typing a misspelled builder name does not produce no rule to make target, it produces whatever Sphinx says about an unknown builder. That is a small ergonomic cost for not having to enumerate builders.
There is also a GH_PAGES_SOURCES variable holding the source directory and the Makefile itself, which is the list of files GitHub Pages needs to publish, and ZIPOPTS defaults to quiet overwrite so unzipping a dataset archive twice does not prompt.
Dataset downloads come with a warning about silent CI breakage
One target handles data, and it opens with a warning that is more interesting than the commands under it.
The note says to make sure a dataset is downloaded into the correct source directory, otherwise CI might silently break. That is the kind of comment written by someone who has seen a build fail for a reason nobody could see.
The commands themselves are ordinary. A Python script runs first and is described as performing structured downloads, which also creates the directories it needs. After that come a wget for the faces archive from the PyTorch download host and an unzip into the beginner tutorial data directory, then another wget for a checkpoint archive that is copied rather than extracted.
A third section prepares the neural style images, and it starts by removing an existing directory with a forced success so a stale copy cannot survive a failed download.
What is worth noticing is the pattern rather than any single file: the data is fetched at build time from URLs outside the repository, the paths are hard-coded per tutorial directory, and the failure mode when a path is wrong is a broken build rather than a clear message. Anyone forking this should read the warning before changing a directory name.
The licence follows upstream, which is the right default
Licensing is stated as a one-liner: the Korean PyTorch tutorials follow the original tutorial's licence, BSD-3-Clause, with the LICENSE file in this repository holding the detail.
Inheriting the upstream licence rather than choosing a new one is the correct default for a translation, and stating it as inherited rather than as chosen removes the question for anyone who wanted to reuse the text.
The rest of the repository administration is thin in the way a documentation fork should be. There is a contributing guide, a translation guide, a licence, a Makefile, the Sphinx sources and the content directories. There is no package manifest and no setup file, because there is nothing to install: the output is a website, built with Sphinx and published to a custom domain.
There are no GitHub releases beyond the tagged documentation snapshots, and the last push is dated 28 July 2026.
One practical note for anyone using this repository as a reference: because the fork tracks one upstream commit, translating from it gives you the text as of that commit, not as of today. For current material, the readme points at the official English tutorials, and for current Korean material, the site is the place to look even if it lags.
Editorial conclusion
tutorials-kr is a good answer to a narrow and real problem, which is that translated framework documentation rots the moment upstream moves. Pinning the upstream commit and tagging the first and last release of each version means a reader on an older PyTorch can still read the Korean text that matched it. Two things to check first: the readme says translations are folded into the site irregularly, so the live site may lag the repository, and the pre-1.0 material is in separate repositories that are no longer being translated. If you want to contribute, reviewing is the role it is shortest of.
Frequently asked questions
What is tutorials-kr and what does it translate?
It is the Korean translation of the official PyTorch tutorials, currently working from PyTorch v2.8 at upstream commit c4d9d93. It is built as a Sphinx site and published at tutorials.pytorch.kr.
How do I read the Korean tutorials for an older PyTorch version?
From v1.0 onwards the versions are not in separate repositories: find the release in this repository's release menu, download the documents and build them locally. A tag named version-base marks the release at the start of that version and version-latest the last release of it.
How can I contribute to tutorials-kr?
Three ways: fix a typo found on the Korean tutorials site, translate a tutorial that has not been translated yet, or review a document that already has a pull request. The reviewing role is marked as the one the project most needs people for, and the contributing guide has the details.
What licence are the Korean PyTorch tutorials under?
BSD-3-Clause, inherited from the original tutorial licence rather than chosen separately. The LICENSE file in the repository holds the full terms.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/pytorchkr-tutorials-kr)Star history chart for your README
[](https://hysenlabs.com/projects/pytorchkr-tutorials-kr)Was this useful?