README.md

metadata

license: apache-2.0
tags:
  - multilingual
  - code
  - programming
  - localization
  - terminology
pretty_name: Legesher Language Canon
language:
  - am
  - ar
  - bn
  - bs
  - ca
  - cs
  - cy
  - da
  - de
  - el
  - en
  - es
  - fa
  - fi
  - fr
  - ga
  - gd
  - gl
  - gu
  - he
  - hi
  - hr
  - hu
  - id
  - it
  - ja
  - kn
  - ko
  - lv
  - mk
  - ml
  - mn
  - ms
  - my
  - nb
  - nl
  - pl
  - pt
  - ro
  - ru
  - si
  - sr
  - su
  - sv
  - te
  - th
  - tr
  - ur
  - uz
  - vi
  - zh
configs:
  - config_name: canon
    data_files:
      - split: train
        path: data/canon.parquet
  - config_name: locales
    data_files:
      - split: train
        path: data/locales.parquet

Legesher language canon

What this Canon is

A centralized standard, created and curated by language communities, for what programming concepts are called in each natural language.

It holds language-specific renderings of programming keywords, builtins and exceptions, and documents where each language stands in that process.

The language-canon builds on legesher/language-corpus, which gathers every rendering anyone has proposed. The corpus is the full record of what has been suggested; the canon is the vocabulary that downstream applications actually use.

Computer science as a global discipline is primarily based in English. A developer who does not read English memorizes shapes rather than reading words and connecting the terminology to related context. The hard part is not the obvious vocabulary. It is the gray areas, where words are borrowed, created, or derived from English (elif and lambda have no obvious answer in any language), and settling them takes people who know a language's culture and the stories woven into it, not only its grammar.

Note: This is a growing dataset. A rendering here is the current answer, not a final one, and each carries a rung recording how much review stands behind it.

Renderings are keyed by content-addressed uids (opt_ + 24 hex) shared with the corpus, so an option_uid resolves there to every alternative proposed for that concept and every endorsement attached to them.

Languages

All 51 languages in this export, with where each one stands on the maturity ladder. The primary-subtag rollup in the frontmatter is what the Hub indexes for search; this is the full list, script subtags and all.

Tag	Language	Native name	Status
am	Amharic	አማርኛ	machine
ar	Arabic	العربية	experimental
bn	Bengali	বাংলা	experimental
bs	Bosnian	Bosanski	machine
ca	Catalan	Català	machine
cs	Czech	Čeština	machine
cy	Welsh	Cymraeg	machine
da	Danish	Dansk	machine
de	German	Deutsch	experimental
el	Greek	Ελληνικά	machine
en	English	English	machine
es	Spanish	Español	experimental
fa	Persian	فارسی	machine
fi	Finnish	Suomi	experimental
fr	French	Français	machine
ga	Irish	Gaeilge	machine
gd	Scottish Gaelic	Gàidhlig	machine
gl	Galician	Galego	machine
gu	Gujarati	ગુજરાતી	experimental
he	Hebrew	עברית	experimental
hi	Hindi	हिन्दी	experimental
hr	Croatian	Hrvatski	machine
hu	Hungarian	Magyar	machine
id	Indonesian	Bahasa Indonesia	experimental
it	Italian	Italiano	experimental
ja	Japanese	日本語	machine
kn	Kannada	ಕನ್ನಡ	machine
ko	Korean	한국어	machine
lv	Latvian	Latviešu	machine
mk	Macedonian	Македонски	machine
ml	Malayalam	മലയാളം	experimental
mn-Cyrl	Mongolian	Монгол	machine
ms	Malay	Bahasa Melayu	machine
my	Burmese	မြန်မာ	machine
nb	Norwegian	Norsk	machine
nl	Dutch	Nederlands	machine
pl	Polish	Polski	machine
pt	Portuguese	Português	experimental
ro	Romanian	Română	machine
ru	Russian	Русский	machine
si	Sinhala	සිංහල	machine
sr-Cyrl	Serbian	Српски	machine
su	Sundanese	Basa Sunda	machine
sv	Swedish	Svenska	machine
te	Telugu	తెలుగు	experimental
th	Thai	ไทย	machine
tr	Turkish	Türkçe	machine
ur	Urdu	اردو	experimental
uz-Latn	Uzbek	O'zbek	machine
vi	Vietnamese	Tiếng Việt	machine
zh-Hans	Chinese	中文	experimental

Today: 36 at machine, 15 at experimental.

The locales config carries the same list with the counts behind each status, so you can sort and filter it rather than read it.

The maturity ladder

Without other global examples to build from, each language's canon starts drafted from machines and prior art. From there native speakers move it up: experimental, then reviewed, then official. Once a language reaches official, its canon becomes a resource for the community it exists to serve.

The rung belongs to the language. Each rendering carries its own so the label stays auditable, but what you should read is the language's status: a language is experimental while it is in progress, and reviewed or official only once every one of its words has got there.

machine

Drafted, not decided. The vocabulary came from a machine draft or from prior art, and nobody has weighed in on any of it yet.

It is not arbitrary. Every rendering clears mechanical checks before it can be selected: it has to be a valid identifier in its language, stable under the Unicode normalization the interpreter applies, and free of collisions with other terms in its vocabulary. What nobody has asked is whether it is the right word — whether a programmer who speaks the language would recognise it.

Treat the whole vocabulary as a proposal.

experimental

Review has begun. At least one word has been chosen or endorsed by a person — a maintainer curating a batch, a native speaker correcting a term, or a contributor putting their name behind a rendering someone else proposed.

Most of the endorsements here predate this dataset: they are the review history of the community that built Legesher's original translations, where contributors voted on renderings word by word. An endorsement attaches to the specific option that was later selected, so it is evidence about this word, not about the concept in general.

A language stays here for as long as the work is under way, however far along it is. The locales config carries the counts if you want to know how far.

reviewed

Every word has been signed off by several native speakers, so the vocabulary reflects agreement rather than one person's preference.

The risk with a single reviewer is not that they are wrong. It is that they are unrepresentative — a regional form, a register that reads formal or childish elsewhere, a loanword one community has absorbed and another rejects. More than one voice is what makes a vocabulary a shared answer.

How many voices, and how varied, is still being defined. We would rather say so than publish a threshold we have not set.

official

Ratified in a convened session, and screened.

Screening is separate from accuracy. A rendering can be a correct translation and still be the wrong word to ship: offensive in another register, unfortunate beside a neighbouring term, or colliding with something it should not. A status of official means both questions were asked of the whole vocabulary and answered by people with standing to answer them.

A language community makes this decision, not a threshold.

Where things stand today

36 at machine, 15 at experimental.

Review has begun in 15 languages, covering 505 of 9,375 renderings. No language has reached reviewed: that needs every word signed off by more than one native speaker, and the threshold for how many voices is still being defined.

We publish at this stage deliberately. A terminology set developed privately and released once it is "ready" has already made every decision without the people it is for. Publishing early makes the current state visible, citable and arguable, while it is still genuinely open.

A language's status here describes its canon. It is not the same as the tier of the language pack that ships from it.

Configs

Two, joined on locale.

canon — one row per decision.

Column	Meaning
concept	Language-neutral concept slug — what the construct does
category	Concept category: control flow, declaration, error handling, …
locale	BCP 47 language tag
option_uid	Content-addressed id of the chosen rendering; joins to the corpus
surface_form	The word itself
status	Rung on the maturity ladder
identity_ok	The rendering matches the English term on purpose
decided_by	Role that made the selection
decided_at	When the selection was made

identity_ok marks a deliberate cognate or loanword — stopiteration in Danish, Lambda in German. It does not mean "not translated yet"; without it a reader could not tell a considered borrowing from an oversight.

locales — one row per language: locale, language_name, native_name, rtl, decisions, a count at each rung, and native_reviewed. Language attributes live only here, so a language is described in exactly one place.

Who decides, and how to take part

decided_by records the role that made a selection, never the person: automated for migrations and batch jobs, maintainer for curation, native-reviewer for a speaker of the language. Contributor identities are not published in this dataset. Endorsements, attribution and withdrawal are handled through legesher/language-corpus, which is the artifact that carries them.

The role answers a different question from the rung, and the two are deliberately independent: status says how much review stands behind a rendering, decided_by says who made the most recent call. A single native reviewer is visible here without moving the rung, because their judgement is real and is not yet a consensus.

This requires collaboration from people who do not just speak a language other than English, but know its culture. If that is you, corrections, endorsements and objections all move a rendering up the ladder, and disagreement is a contribution rather than an obstacle.

Licence and citation

Released under Apache-2.0. Legesher holds the compilation right in this dataset and licenses it on those terms; the individual renderings arrive from the corpus, whose own licence and attribution notices govern the contributed parts.

@misc{legesher_language_canon,
  title  = {Legesher Language Canon},
  author = {Legesher},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/legesher/language-canon}}
}