arabic-lint

Finds Arabic text that was corrupted before it was stored — in your JSON, your localisation files, your database exports, your source code.

pip install arabic-lint
arabic-lint ./src

src/strings.json:3:20: 21 Arabic presentation forms stored  [UNSAFE TO AUTO-FIX]
    found     : ﺓﺪﺤﺘﻤﻟﺍ ﺔﻴﺑﺮﻌﻟﺍ ﺕﺍﺭﺎﻣﻹﺍ
    would be  : اإلمارات العربية المتحدة
    contains a lam-alef ligature; NFKC decomposition reorders the pair, so this
    recovery is wrong even though it looks like Arabic

3 corrupted span(s) in 1 file(s); 1 cannot be auto-fixed safely.

Exit code 1 when anything is found, so it drops into CI unchanged. MIT. Zero dependencies. Python 3.9+.

Does this actually happen in the wild?

Yes, and here is the measurement rather than the assertion.

syamjithnk/arabic-corpus-audit — 341 public Arabic datasets on the Hugging Face Hub, 276 readable, 26,318 rows, 119,517 text fields, scanned with this tool.

• 1 dataset carries presentation forms in every sampled label. Yousefmd/arabic_ocr_dataset: 1,000 of 1,000 sampled labels, 3 to 50 presentation forms per label (median 24), INITIAL/MEDIAL/FINAL/ISOLATED forms together. It is an OCR set, so the affected field is the ground-truth label — a model trained on it learns to emit glyph forms that will not compare equal to ordinary Arabic. Only the shaping step ran there, not bidi: the words are in logical order and plain NFKC recovers them. An earlier version of this README said visual order; that was wrong and is corrected in the audit's own correction log. (Its scale, stated plainly: 17 downloads. This proves the mechanism reaches training data; it is not evidence that widely-used corpora are affected.)
• 20 more had single stray presentation forms, no reshaped runs. Still wrong, since a word becomes [UNK] on a tokenizer that does not normalise Unicode. Measured on four real vocabularies: AraBERT v02 and mBERT replace the word with [UNK] 6 times out of 6, so every character of meaning is discarded before the model sees it. One corrupted letter out of four is enough. XLM-R normalises and is unaffected. A smaller defect than a reshaped corpus, but not a cosmetic one.
• Everything else was clean.

The audit ships its own scanner, so the numbers are re-derivable rather than trusted. It also documents a false positive this tool used to produce against Islamic heritage text, which is worth reading before you point any such tool at someone else's corpus.

Confirmed in the wild. Four bug reports were filed from the source check below. Two were confirmed and closed as fixed on 12 September 2026, in whiteout-project/bot#110 and its sibling kingshot-project/Kingshot-Discord-Bot#28, by the maintainer of both. The maintainer also found the same pattern in a second code path neither report mentioned. Two reports, ComfyUI-PersianText#3 and Diwan#3, are still open.

Severity: one pasted glyph is not a destroyed corpus

The audit settled this empirically. Across 276 public Arabic datasets, 361 of 363 findings were a single stray presentation form in otherwise correct text, and exactly one dataset carried long runs. Those are different problems:

severity	forms	what it means	what to do
stray	1	pasted from a PDF, or OCR residue	fix the character
partial	2–4	a fragment, or a short pass through the recipe	check where the text came from
reshaped	5+	a shaping pass ran before this was stored	audit the pipeline, not the file

Reporting them identically means a team with one stray glyph in ten thousand rows gets the same alarm as a team whose corpus was destroyed, and then they switch the alarm off.

arabic-lint . --min-severity reshaped     # fail CI only on pipeline damage

Run it on commit, or in CI

Corruption at rest is cheap to catch and expensive to find later, because by the time anyone notices, the pipeline that produced it has written the same text into several other places. Both integrations below default to the reshaped gate for the reason in the table above.

pre-commit

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/Syamjith-NK/arabic-lint
    rev: ""        # run `pre-commit autoupdate` to fill in the latest tag
    hooks:
      - id: arabic-lint

pre-commit install
pre-commit run arabic-lint --all-files

The hook installs into pre-commit's own isolated environment, and because the package has no dependencies there is nothing to resolve. Use id: arabic-lint-strict instead if you want a single stray presentation form to fail the commit too.

GitHub Actions

# .github/workflows/arabic-lint.yml
name: arabic-lint
on: [push, pull_request]
jobs:
  arabic-lint:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: Syamjith-NK/arabic-lint@main    # or pin to a tag

Inputs, all optional: path (default .), min-severity (default reshaped), args for anything else the CLI takes, and version to install a published release from PyPI instead of the code at the ref you pinned.

      - uses: Syamjith-NK/arabic-lint@main
        with:
          path: locales
          min-severity: stray
          args: --exclude vendor --exclude fixtures

The job fails on findings, because that is what exit code 1 is for. Two things worth knowing before you turn either of these on:

• The severity gate applies to stored text. A source finding, the reshape+bidi recipe feeding a renderer that already shapes, is not graded by severity and is always reported. That check is quiet by design and does not fire on the thousands of files where the recipe is correct, so it is not the thing that will flood you.
• --fix is not wired into either integration, and should not be. Which repair is right depends on whether you control your dependency floor, which is not in the source file: see which fix is correct depends on your dependency floor.

Adopting it on a codebase that already has findings

The exit code is binary, so on an existing repository the check fails on day one and the usual outcome is that nobody turns it on. Record what is already there, then gate on what gets added next:

arabic-lint . --write-baseline .arabic-lint-baseline.json   # what exists today
arabic-lint . --baseline .arabic-lint-baseline.json         # fail only on what is new

no NEW findings - 412 file(s) scanned.
2 finding(s) suppressed by the baseline.

Commit the baseline. Four things about it are deliberate:

• Entries are keyed on content, not on line numbers. Adding an import at the top of a file does not resurrect an entry, and does not hide a new one either.
• The grade is part of the key. A stray glyph that later grows into a reshaped run is a new problem in the same file, and is reported.
• Identical findings are counted, not collapsed. One baselined U+200E excuses one, not any number of them.
• Suppressions are always printed, and a baselined finding that has since been fixed is counted as no longer present. Silence about what a baseline hides is how these files quietly become permanent; prune them by regenerating.

Paths inside the file are relative to the baseline itself, so it means the same thing whether CI runs from the repository root or a subdirectory. Samples are stored as escaped codepoints (U+FEE3 U+FE8E) rather than raw text — the file would otherwise be full of presentation forms and the next scan would report every one of them.

--min-severity still applies while writing, so --write-baseline --min-severity reshaped records only pipeline damage and leaves stray glyphs reportable.

Which case are you in?

The findings below all depend on what draws your text, and that is not knowable by reading code: it depends on your installed matplotlib version and on whether your Pillow was built with Raqm, which is a property of the wheel rather than the version.

arabic-lint --doctor

  matplotlib 3.11.0
      pre-shaping BREAKS text here
      3.11 and later shape text with libraqm and apply bidi themselves

  Pillow 12.2.0
      pre-shaping BREAKS text here
      this build has Raqm, so ImageDraw.text() shapes and reorders for you. Note
      this is a property of the BUILD: the same version on another machine can
      answer differently

Verdict: on matplotlib and Pillow, remove the reshape/bidi step and pass the
logical string straight through. Leaving it in reverses the text, silently.
That is THIS environment. Code installed on machines you do not control
should gate on the renderer version instead of removing the step.

If one renderer shapes and the other does not, it says so, because then a single shared helper cannot be correct for both.

--doctor answers for the machine it runs on. If your code is installed on other people's machines, that is a different question, and it changes the fix: see which fix is correct depends on your dependency floor.

What counts as corruption, and what does not

Arabic Presentation Forms-A is interleaved: positional glyph forms and ordinary semantic characters share the same block. The ornate parentheses ﴾ ﴿ that enclose a Quranic quotation live at U+FD3E/U+FD3F, in the middle of it, and honorific ligatures like ﵀ sit just above them. Those are characters people type on purpose.

Treating the whole block as a corruption signal reports Islamic heritage text as broken. Measured against public corpora on Hugging Face on 9 September 2026, the naive rule flagged 35.6% of one heritage OCR corpus and 5.5% of another. Every hit was a Quranic quotation mark or an honorific. With the block classified properly, both read 0%.

So the classification is derived rather than tabulated: a contextual shaping artefact is exactly a codepoint Unicode names ... ISOLATED/INITIAL/MEDIAL/FINAL FORM. Anything else in the block is deliberate. New Unicode additions classify themselves.

Three checks

Stored corruption — Arabic presentation forms that were already written to disk.

Bidi controls — the twelve invisible directional characters, when they survive into Arabic text or into source. See invisible bidi controls below.

Source risk (Python files and notebook code cells) — the reshape+bidi recipe feeding a renderer that already shapes, which corrupts at render time before anything is stored:

cogs/bear_track.py:715:16: pre-shaped text passed to matplotlib  [RENDERS REVERSED]
    return _bidi_get_display(_arabic_reshaper.reshape(text))
    matplotlib >= 3.11 shapes text with libraqm and applies bidi itself, so this
    reorders an already-reordered string and the text renders reversed

The recipe appears in 3,168 indexed Python files on GitHub (measured 2026-09-08; re-measured 3,128 on 2026-09-13, because the count drifts as GitHub reindexes, so re-run the search rather than trusting the number). Flagging all of them would be worthless, because whether it is a bug depends entirely on what draws the text:

renderer	shapes and reorders?	verdict
matplotlib >= 3.11	yes	pre-shaping reverses the text
Pillow built with Raqm	yes	pre-shaping reverses the text
Pillow without Raqm	no	pre-shaping is required
ReportLab, fpdf	no	pre-shaping is required
print() to a terminal	terminal-dependent	not this tool's call

So the source check stays silent unless a shaping renderer is imported and something in the file actually draws with it. It also follows import aliases, follows the recipe when reshape() and get_display() are on separate lines, names the renderer that actually draws rather than the first one imported, and ignores a pre-shaping helper that a script never calls.

Validated against six real repositories found by code search: it flags the three that are genuinely broken, and stays silent on a ReportLab project, a dead helper in an unrelated benchmark, and a script that only prints to a terminal. Pass --no-source to turn it off.

A finding here says the recipe will corrupt text on a shaping renderer. It does not say which repair is right for your project, and the report text overstates the case when it suggests simply removing the call. Removing it is correct only if you control which version of the renderer your code runs against: which fix is correct depends on your dependency floor.

--fix, and why it exists here but not for stored text

The same package refuses to repair one kind of damage and offers to repair the other, which sounds inconsistent until you look at what each one is.

Stored corruption is not safely repairable. Undoing it round-trips exactly until the text contains a lam-alef ligature, and then السلام comes back as السالم: a real word, a different word, one that survives a proofread. So the stored check reports and never rewrites.

Source is the opposite. Where the whole recipe sits inside one expression, deleting it is exactly correct:

- return get_display(arabic_reshaper.reshape(text))
+ return text

arabic-lint . --fix makes that edit in place, then refuses to write if the result would not parse.

It does not touch the split form:

reshaped  = arabic_reshaper.reshape(segment)
processed = get_display(reshaped)        # NOT auto-fixable

Rewriting the second line to processed = reshaped would remove the explicit reordering but leave the shaping applied. A shaping renderer still reorders that text, so it can look correct while differing on ligatures. On Pillow without Raqm, it renders in the wrong order because the deleted call was doing the reordering. Both lines must be handled together, and which other code reads reshaped is not knowable from that expression. Pass the original string to a shaping renderer, or gate both steps on the renderer's shaping support. The tool leaves the file alone.

Which fix is correct depends on your dependency floor

Deleting the call is not always the right repair, and this tool cannot tell which case you are in, because the answer is not in the code. It depends on whether your project controls the version of the renderer it runs against.

If you set the floor, delete the call. A repo that pins matplotlib>=3.11, an application shipped with a lockfile, or anything whose supported versions are declared in CI knows that the renderer shapes. That is the rewrite --fix performs.

If you cannot force an upgrade, gate on the version instead. A distributed application whose users update on their own schedule, or a library installed next to whatever the user already has, will keep running on older renderers after your fix ships. For those installs a bare removal leaves no shaping at all, which trades reversed text for disjointed text: a different bug, not a fix.

import matplotlib
from packaging.version import Version

if Version(matplotlib.__version__) >= Version("3.11"):
    label = text                                    # matplotlib shapes and reorders
else:
    label = get_display(arabic_reshaper.reshape(text))

This is not a hypothetical. It is the fix the maintainer of whiteout-project/bot and kingshot-project/Kingshot-Discord-Bot applied in September 2026, in preference to the removal these reports proposed, and the reason is one no static analyser can reach: their updater installs missing packages but never bumps existing ones, so a bare removal would have left every existing 3.10 install with no shaping at all.

So treat --fix as correct for code whose renderer version you pin, and read a source finding as this will corrupt text on a shaping renderer rather than as delete this line. Note also that --doctor reports the machine it runs on, which is the right answer for a repo you deploy and the wrong one for software other people install.

Why this exists: the recipe appears across the indexed Python files counted in Two checks above, and on matplotlib 3.11 it now renders Arabic backwards with no error at all. The Arabic fix everyone recommends is now the bug — the measurements, and what happened when it was filed against Pillow and matplotlib.

Check it yourself

Do not take the claim on trust — verify_mpl311.py re-measures it on your machine:

pip install matplotlib arabic-reshaper python-bidi
python3 verify_mpl311.py

It renders each test string twice, once as typed and once through the workaround, and reports the mean absolute pixel difference against a same-string control. On matplotlib 3.11.0 with Pillow 12.2.0 (Raqm enabled) all five strings render differently, against a control of exactly 0.000.

The pixel numbers depend on your font and size, so treat them as a signal rather than a constant. The evidence that needs no renderer at all is the codepoint count the script also prints:

emirates    8 →  7 codepoints · 7 in the Arabic Presentation Forms block  ← lam-alef decomposed

الإمارات loses a codepoint because lam-alef is one codepoint that decomposes into two, and every output character lands in a block that typed Arabic never contains.

What it actually detects

The most widely copied recipe for "making Arabic work" in Python is:

text = get_display(arabic_reshaper.reshape(text))

Those two calls do the job a text engine is supposed to do: substitute each letter for its contextual presentation form, and reorder the string into visual order. If your renderer already does complex text layout — Pillow with Raqm, matplotlib, any browser — the work happens twice and the output is wrong.

The real damage is when that string gets written back: to a config, an export, a translation file. Now the corruption is at rest and every downstream reader inherits it. It renders as clean-looking Arabic, so nobody who does not read the script will ever notice.

The signature is unambiguous: Arabic presentation form codepoints in stored text — all of Forms-B (U+FE70–U+FEFF), plus the positional forms in Forms-A (U+FB50–U+FDFF). Those exist for legacy-encoding compatibility; correctly authored modern Arabic never contains them.

The Arabic word ligatures ﷲ ﷺ ﷻ ﷽ (U+FDF0–U+FDFD) are deliberately excluded — people type those on purpose, and flagging them would mark correct religious and formal text as corrupt.

Forms-A matters more than it looks. Persian and Urdu share most of their alphabet with Arabic, so most reshaped Persian already trips Forms-B — but it was being under-counted, and words built only from Persian-specific letters (گچ, چپ, گپ, پژ, پی, and a bare Farsi yeh) have no Forms-B mapping at all and were missed outright.

Why it reports instead of fixing

You would think you could just undo it: NFKC maps every presentation form back to its base letter, and reversing undoes the visual reordering. That round-trips exactly — until the span contains a lam-alef ligature.

A lam-alef ligature (لا, لأ, لإ, لآ) is one codepoint standing for two letters. NFKC expands it in logical order while the text around it is still in visual order, so the pair comes out reversed relative to its neighbours:

original	naive "fix"
الإمارات	اإلمارات	not a word
السلام	السالم	a real but different word

That second row is the whole reason this tool exists rather than a sed command. The output is still pronounceable Arabic, so it survives a proofread — and the definite article followed by alef is one of the most common sequences in the language, so this is not a corner case.

arabic-lint shows you the candidate recovery and tells you when it is unsafe. It never rewrites stored text: --fix applies to source findings only, and refuses even there when the two halves of the recipe are split across separate lines.

One important qualifier. All of the above assumes both calls ran. reshape() on its own, with no get_display(), is common in the wild: it leaves presentation forms in logical order, and then NFKC alone is a complete and safe repair, lam-alef included. So check word order before you assume the harder case. If the string's words come back readable under plain NFKC, bidi never ran and there is nothing to reverse. Getting this backwards means telling someone their data is harder to repair than it is.

Verification

• The test suite runs with no runtime dependency on arabic_reshaper or python-bidi: fixtures are recorded from a real run of both, and the two tests that do call them skip rather than fail when they are absent.
• Block boundaries were measured, not assumed: over a wide Arabic sample, arabic_reshaper 3.0.0 emits 53 distinct codepoints from Presentation Forms-B and never emits U+FEFF.
• Validated against 3,826 real files — the false positives that scan found are now regression tests:
  • U+FEFF sits inside Forms-B but is the byte order mark. Excluded.
  • Presentation Forms-A is deliberately not a signal. The reshaper emits exactly one codepoint from it (U+FDF2, the Allah ligature), and that character — like ﷺ U+FDFA, ﷻ U+FDFB and ﷽ U+FDFD — is used intentionally in ordinary Arabic writing. Treating the block as corruption flags correct religious and formal text.

Invisible bidi controls

Twelve characters in Unicode carry direction and nothing else. They are zero-width, they survive a proofread, and almost nothing handles all twelve the same way.

The blind spot is real and measured. nmt_nfkc, the default normalizer for SentencePiece training, maps U+200E LEFT-TO-RIGHT MARK and U+200F RIGHT-TO-LEFT MARK to a space and leaves the other ten alone — including U+061C ARABIC LETTER MARK, which does for Arabic-script runs exactly what those two do everywhere else. On google/mt5-base, a phrase goes from 5 pieces to 7 with an ALM in it and is unchanged with an RLM; google/mt5-base also carries U+061C as its own vocabulary token. Same visible text, different token sequence, decided by which pipeline saw it.

Filed upstream as google/sentencepiece#1331, and closed as fixed on 22 September 2026 by Taku Kudo, SentencePiece's author, who mapped U+061C to a space in nmt_nfkc beside LRM and RLM and added a regression test covering all three. The report was the contribution; the fix is his. It is in the repository and not in a release: v0.2.2 is the newest and predates it, so the normalizer you install today still passes ALM through, and a model that already exists keeps the precompiled normalizer it was trained with. That is the residue band in the table below, and the reason this check is here.

An unterminated scope is a different problem with a different weight:

config.py:14:31: U+202E RIGHT-TO-LEFT OVERRIDE [unpaired]
    kind      : override (bidi class RLO), offset 402
    balance   : never closed, so its scope runs to the end of the paragraph
    the directional scope is not paired. An unclosed embedding or override applies to
    everything after it to the end of the paragraph, so what is displayed is not the
    order the text is stored in. Close it, or remove it.

That is the Trojan Source class: the line displays in an order it is not stored in. A balanced RLE … PDF is ordinary directional markup and is graded scoped, not flagged as a defect.

risk	what it is	gated by
residue	a lone ALM / LRM / RLM	--min-severity stray (default)
scoped	an explicit run, opened and closed	--min-severity partial
unpaired	a scope with no terminator, or a terminator with no scope	--min-severity reshaped

One --min-severity floor gates both ladders by position, so --min-severity reshaped narrows stored findings to pipeline damage and controls to unpaired ones. The words are not shared on purpose: a stored reshaped means a shaping pass ran, and printing that beside an unterminated override would be false.

What stays quiet, and why. A mark or a closed scope is reported only when its paragraph actually contains Arabic — otherwise this would fire on every correctly marked-up Hebrew document on earth. An unpaired control is reported whatever the script, because an unterminated override reorders whatever follows it, and the Trojan Source case lands in source files with no Arabic in them at all.

The nine explicit formatting characters are derived from their Unicode bidi class, not tabulated, so additions classify themselves — the same rule as the presentation forms. The three directional marks are named, because the property that identifies them is Bidi_Control, which stdlib unicodedata does not expose. The obvious substitute is wrong and the suite proves it: Cf + a strong bidi class matches 23 characters on Unicode 16.0, among them the Syriac abbreviation mark, two Kaithi number signs and sixteen Egyptian hieroglyph joiners — characters people type on purpose. Flagging those would be the ornate-parentheses mistake in a different block.

--fix never touches a control. Whether one belongs in a document is a question about the document, not about the character. Pass --no-controls to turn the check off.

Jupyter notebooks

.ipynb is scanned like any other file, and findings name the cell:

analysis.ipynb:cell2:5:11: pre-shaped text passed to matplotlib  [RENDERS REVERSED]
    plt.title(get_display(arabic_reshaper.reshape('مرحبا بالعالم')))
analysis.ipynb:cell3:1:10: 4 Arabic presentation forms stored [partial]
analysis.ipynb:cell3:1:1 (output): 4 Arabic presentation forms stored [partial]

Notebooks matter more than their share of the code: the recipe is copied between notebooks far more than between modules, because it is the thing someone pastes from an answer to make one chart work.

Three things worth knowing:

• The same corruption is reported twice — once in the cell source and once in the committed output — because a notebook stores both. The (output) label is there because you cannot fix an output by editing it; fix the code above and re-run.
• Cells that are not Python are fine. %matplotlib inline and !pip install cost a quiet skip, not a crash and not a false finding.
• --fix will not rewrite a notebook. Changing one cell means re-dumping the whole JSON, which reformats every untouched cell and buries a one-line fix in a whole-file diff. It tells you which cell to edit instead.

Reading needs no new dependency — a notebook is JSON, so nbformat is not required and the tool still installs with nothing behind it.

Known limits

• Balance is computed per paragraph, which is what the bidirectional algorithm does, so an opener on one line and its terminator on the next are reported as two unpaired controls. That is correct — the scope really did end at the paragraph — but it will surprise anyone who wrote them as a pair.
• The control check answers "is this scope closed", not "is this document's direction right". It does not resolve embedding levels, and it cannot tell you whether a correctly balanced override was a good idea.
• A document whose only Arabic is a standalone Allah ligature is missed. That is the deliberate trade above; any corrupted phrase around it still trips Forms-B.
• The recovery direction assumes bidi was applied. Text that was reshaped but not reordered recovers reversed. The tool shows you the candidate so you can see which case you have; it does not guess.
• The stored check only sees corruption that is already written down. The source check is what looks ahead at code that will create some, and --doctor is what answers it for the environment actually doing the rendering.
• Notebook image and HTML outputs are not scanned. A base64 PNG cannot contain the Arabic this looks for, and reading it would make every notebook with a chart slow for nothing. Text, text/plain and tracebacks are scanned.
• Pure reordering is invisible to it, whatever produced it. Reshaping leaves codepoints that authored Arabic never contains, so it is detectable; reordering leaves the same codepoints in a different order, so nothing here can flag it. A live example: pypdf classified the Arabic-Indic digits U+0660–U+0669 as right-to-left, so extraction returned ١٢٣٤ as ٤٣٢١ — valid Arabic digits in the wrong order, which this tool reads as clean. Fixed upstream in pypdf — two lines of library change plus a test, merged 14 September 2026 — and released in pypdf 6.19.0 on 16 September 2026. So the version now answers half the question: below 6.19.0 the digits come out reversed. It does not answer the other half, because text extracted by an older pypdf is already wrong and scanning it will not reveal that. Check the digit order against the source document.
• It cannot tell whether you control your dependency floor, so it cannot tell you whether to delete the pre-shaping call or gate it on the renderer version. That distinction lives in your packaging and your users' upgrade path, not in the source file. See which fix is correct depends on your dependency floor.

Contributing

Renderer knowledge, false-positive reports and test cases taken from real repositories are all wanted. CONTRIBUTING.md has the two rules that matter: no runtime dependencies, ever, and validate against real code rather than invented snippets. The open issues are scoped so a stranger can start on one, and the ones labelled good first issue genuinely are.

A false positive is the most valuable report this project can get. The whole design is built around staying quiet.

Related

Part of a series measuring where Arabic silently breaks in software. See also arabic-tts-frontend — numerals, dates and currency converted to spoken Arabic before synthesis.

The measurement behind this tool

The severity bands here are not a guess. They come from sweeping 341 public Arabic datasets on the Hugging Face Hub: 361 of 363 findings across 276 readable datasets were a single stray presentation form, and exactly one dataset carried long reshaped runs. That is why one pasted glyph and a destroyed corpus are no longer reported identically.

The audit, its full per-dataset results and every script are archived and citable:

• DOI: 10.5281/zenodo.22733934
• Dataset: syamjithnk/arabic-corpus-audit