sukhrobnurali/uzbek-legal-corpus

Uzbek Legal Corpus (full, uz-Latin)

This is a research corpus of published Uzbek legislation — a pinned consolidated snapshot, not a live legal database. It is NOT legal advice and NOT an authoritative legal source. The authoritative source is lex.uz. Always verify the current, in-force text there before relying on any provision.

A cleaned, article-structured corpus of Uzbek (Latin-script) national legislation of every type — laws, codes, the Constitution, presidential decrees and orders, cabinet resolutions, and ministerial orders — sourced from Lex.uz, the National Database of Legislation of Uzbekistan (maintained by the "Adolat" National Legal Information Center under the Ministry of Justice). It covers 24,267 acts / 54,173 article rows spanning 1991–2026, all in native Uzbek Latin script.

This is the full single-language companion to sukhrobnurali/uzbek-legal-corpus-v1 — a small, curated, trilingual showcase of just the 20 codes + the Constitution. Where v1 is depth on 21 foundational acts in three scripts, this dataset is breadth: the whole uz-Latin act set, one script, ~1,200× more distinct acts.

---

Motivation

Uzbek is a low-resource language for NLP, and Uzbek legal text is harder still: it is deeply hierarchical (document → part → chapter → article), and the source HTML mixes the legal text with editorial chrome, amendment notes, and classifier metadata. Most downstream uses — retrieval, RAG, fine-tuning, fill-mask pretraining — want the article as the unit, with its document context attached and the noise removed. That structuring is exactly what is missing from the freely available data.

This corpus was built as the data layer behind AI Huquqshunos, a legal-assistant project for Uzbek users, and then open-sourced as a dataset in its own right.

Where this sits relative to what already exists

To be clear and honest: this is not the first or only Uzbek legal dataset on the Hub, and it is not the first to scrape Lex.uz. Several already exist, for example:

Dataset	Approx. size	Notes
AsrorAsr/lexuz-legal-documents	~27k docs	metadata only, no full text, no card, no license
ninetydevuz/lex-uz-legal-text-scaping	~10k docs	full text, Apache-2.0, but empty README and uncleaned page dumps
Sohiba01/uzbek-legal-ir	~52k rows	no dataset card
AsrorAsr/uzbek-legal-dataset	~1.2k rows	Criminal Code only
Mehriddin1997/lex-uzbek-laws	307 docs	the one with a genuine dataset card

What this dataset adds, stated plainly, is engineering and documentation quality at full scale:

• Cleaned, normalized text — UI chrome and button bars stripped, whitespace collapsed, Unicode NFC-normalized, not raw page dumps.
• Explicit article-level hierarchy — every row knows its part, chapter, and article_number; amendment notes and editorial cross-references are routed to their own fields instead of being inlined into the article body.
• A reproducible, self-verifying enumeration of the whole uz-Latin act set (below), with an honest coverage account.
• Quality flags shipped, not silently dropped.
• A documented sourcing methodology and a clear license with a takedown contact.

If you need raw coverage including Cyrillic/Russian today, one of the larger scrapes above may suit you better. If you need clean, structured, documented uz-Latin text, that is what this is for.

---

Coverage — read this before using

lex.uz indexes each act separately per language: a given act may exist as a uz-Latin document, a uz-Cyrillic document, a Russian document, and/or an English one, each with its own id. uz-Latin transliteration is not universal. Empirically, only about half of the acts on lex.uz have a uz-Latin version at all — the rest are published only in uz-Cyrillic and/or Russian.

Concretely:

• 27,488 distinct uz-Latin acts were enumerated across 1991–2026 (the verified uz-Latin universe; see method below). For comparison, lex.uz holds roughly ~54k acts in uz-Cyrillic — so a uz-Latin-only corpus is inherently about half the size of the full act count. If you need maximum act coverage, uz-Cyrillic is the more complete script.
• Of those 27,488, 3,221 had an empty uz-Latin page (a title-only stub with no server-rendered body — typically acts whose text lives only in the Cyrillic/Russian version) and are excluded.
• The result is 24,267 acts with usable Latin text → 54,173 article rows.

This is a deliberate single-language scope, documented honestly rather than inflated.

---

Dataset structure

Two configurations. articles is the default.

articles (default) — one row per article

The natural unit for retrieval and training. Document-level metadata is denormalized onto every article row so a single row is self-describing.

Column	Type	Description
id	string	Row id, {doc_id}-a{article_number} (or {doc_id}-full for an unstructured act).
act_group_id	string	The act key, abs(doc_id) as a string.
doc_id	int	Lex.uz uz-Latin document id (negative).
doc_title	string	Title of the act.
doc_type	string	law, resolution, decree, order, code, constitution, or other.
doc_number	string	Official act number where the page exposes it (often empty).
doc_date	string	Adoption date (ISO), from the page.
version_date	string	The pinned consolidated ONDATE at which this text was current (see Limitations).
status	string	In-force / repealed, best-effort.
part	string	Part / section heading (BOʻLIM), if any.
chapter	string	Chapter heading (-bob), if any.
article_number	string	Article number, e.g. 12 or 12-1.
article_title	string	Article heading text.
article_text	string	The article body, cleaned.
amendment_note	string	Editorial amendment annotations, routed to their own field.
cross_references	string	Editorial cross-reference / comment annotations, routed to their own field.
language	string	Always uz.
script	string	Always latin.
okoz_codes	list[string]	OKOZ classifier codes (reliable).
tsz_codes	list[string]	TSZ classifier codes (best-effort, known-incomplete — see Limitations).
source_url	string	https://lex.uz/uz/docs/{doc_id}.
n_tokens	int	Whitespace token count of article_text.
quality_flag	string	"" = clean; otherwise too-short, parse-uncertain, or no-articles.

documents — one row per act

The same acts at whole-document granularity, for users who want full context rather than isolated articles. Columns: doc_id, act_group_id, doc_title, doc_type, language, script, version_date, article_count, and full_text (articles concatenated in order).

The hierarchy

document (act)
└── part        (BOʻLIM)
    └── chapter (-bob)
        └── article  ← one row in the `articles` config

part and chapter are carried forward as context onto each article. Most acts (decrees, resolutions, short laws) have no internal article structure and are stored as a single flat-text row ({doc_id}-full, flag no-articles); codes and longer laws split into many articles.

---

Language & script

All text is fetched directly from lex.uz — there is no synthetic transliteration. uz-Latin documents are identified by their negative document id (the uz-Latin id is the negation of the uz-Cyrillic id); only native uz-Latin documents are included.

Text is Unicode NFC-normalized. The Uzbek modifier letters Oʻ / gʻ and the apostrophe ʼ (U+02BB / U+02BC) are preserved as-is — not folded to ASCII — because that fold is lossy for Uzbek orthography. A negligible number of rows (~0.03%) contain short quoted Russian passages embedded in otherwise-Uzbek articles.

---

How it was built

A sequence of small, resumable stages; each logs drop counts so the funnel is auditable end to end.

1. sourcing_check — fetch and evaluate robots.txt; record terms/copyright findings. The gate must pass before anything is crawled.
2. enumerate — discover the uz-Latin act set (below) into a manifest; resumable and incremental.
3. scrape — fetch each uz-Latin page with httpx and gzip-cache it to data/raw_html/. Resumable; cached pages are never re-fetched.
4. parse — with selectolax, split each document on its article-heading CSS class, carry part/chapter context onto each article, and route amendment notes and editorial cross-references into their own fields. UI button-bar chrome is stripped before any text is read. Doc-level metadata (dates, status, OKOZ/TSZ) is read from the page itself.
5. clean — NFC-normalize and collapse whitespace; assign quality flags; drop empty-body rows (see dedup note).
6. build_dataset — assemble the two configs and (optionally) push_to_hub.

Enumeration — how the uz-Latin act set was found, and verified

lex.uz has no sitemap and no bulk API, and its search is an ASP.NET WebForms app. The bare year-filter listing is dominated by non-Latin documents, but a Latin-script query term returns native uz-Latin (negative-id) documents. So, per year (1991→2026), the pipeline unions the native-Latin hits of a few ultra-common Uzbek-Latin words (va, toʻgʻrisida, …), keeping only negative ids, and stops once the union stabilizes (an added term contributes <0.5% new acts). A single common word covers ~99% of a year's uz-Latin acts; the union closes the rest.

Completeness was checked three independent ways and agrees: every year converged (0 unstable years), and adding diverse out-of-set terms (tasdiqlash, kiritish, chora, tartib, …) yields zero new acts. The manifest (data/manifest.jsonl) is committed for reproducibility.

Crawl rate (honest disclosure)

robots.txt states Crawl-delay: 20 with no Disallow. This corpus is far larger than the v1 showcase, so rather than a fixed 20 s delay it uses a conservative AIMD controller: concurrency capped at 4, 0.3 s launch spacing, additive-increase after clean responses, halve-on-error, and it honors Retry-After. In practice this sustained **3–4 requests/second** against a single host; the full ~24k-page crawl took ~3–4 hours, runs unattended, and is fully resumable from the gzip cache (a killed run re-fetches nothing already held). The crawler sends a descriptive User-Agent with a contact address (uzbek-legal-corpus-crawler/1.0 (+sukhrobnurali@gmail.com; portfolio research dataset)). This is brisk but bounded and polite; it is disclosed here rather than hidden.

On duplicates

No cross-document text de-duplication is applied, by design. In legislation, identical text across acts is normal and legitimate: a shared "this law enters into force on publication" article belongs to every law that contains it, and templated decrees (e.g. treaty-ratification laws that differ only by date and counterpart) are separate, individually-citable legal instruments. Collapsing them would delete real acts. Every distinct (act, article) is therefore kept; only genuinely empty-body rows are dropped.

Quality flags

quality_flag is shipped, not hard-cut:

Flag	Meaning
""	Clean.
too-short	Below the minimum token threshold.
parse-uncertain	Structure extracted with low confidence.
no-articles	The act did not split into articles; stored as flat text (most decrees/resolutions).

Corpus statistics

Pipeline funnel:

Stage	Count
uz-Latin acts enumerated (1991–2026)	27,488
Excluded — empty-body / unavailable Latin page	3,221
Documents (final)	24,267
Article rows (final)	54,173

Configs: articles 54,173 rows · documents 24,267 rows.

By document type (documents):

Type	Documents
Resolution (qaror)	16,735
Order (buyruq / farmoyish)	3,232
Decree (farmon)	2,000
Law (qonun)	1,717
Code (kodeks)	259
Constitution	113
Other	211

Structure: 863 acts split into multiple articles (codes, longer laws); 23,404 are single flat-text acts (decrees, resolutions, short laws).

Quality flags (articles): clean 29,198 · no-articles 23,306 · parse-uncertain 1,658 · too-short 11.

Volume: ~47.3M article tokens (whitespace count); mean 873 tokens/article; max 226,647.

---

Licensing & takedown

• Code, pipeline, schema, and annotations (the structuring, the quality flags, the cross-reference/amendment routing): Apache-2.0.
• The legislative text itself: public-domain-exempt under Article 8 of the Law on Copyright and Related Rights (LRU-42/2006), which excludes official documents and their official translations from copyright. Wikimedia Commons treats Uzbek legal texts as PD-UZ-exempt. Reproduced here with attribution to lex.uz as the source.

This dataset reuses the legal text. It does not redistribute lex.uz branding or wholesale-copy lex.uz's proprietary editorial/compilation layer; the on-page editorial annotations that are kept (amendment notes, cross-references) are isolated in their own fields and attributed.

Sourcing checks. The /agreement page on lex.uz is in practice a privacy policy and imposes no restriction on automated access to or reuse of the legislative texts. The national open-data portal data.egov.uz publishes legislation registries/metadata under CC BY-SA 4.0 but not the full normative text; the full text lives only on lex.uz.

Takedown / corrections: sukhrobnurali@gmail.com. If you are a rights holder or the source maintainer and want something changed or removed, contact me and I will act on it.

---

Limitations

• uz-Latin only, ~half the act universe. About 50% of lex.uz acts have no uz-Latin version; those exist only in uz-Cyrillic/Russian and are out of scope here. This is a script-coverage limit of the source, not a crawl gap.
• Pinned snapshot, not a live or amendment-history view. Each act is captured at a single consolidated version_date (ONDATE). The dataset is not updated as the law changes. Always check lex.uz for the current text.
• status is best-effort, extracted heuristically; may be stale or wrong for a given row.
• tsz_codes are known-incomplete (the TSZ classifier embeds code and label in nested elements). OKOZ codes are reliable.
• doc_number is often empty — it is not reliably exposed in the uz-Latin page HTML.
• A handful of documents (≈3) with non-standard markup did not parse and are excluded; the vast majority of excluded docs are genuinely empty-body Latin stubs.
• Not legal advice, not authoritative. See the disclaimer at the top.

---

Intended uses

• Information retrieval and RAG over Uzbek legislation (article-level units with document context).
• Fine-tuning and instruction-tuning of Uzbek legal-assistant models.
• Fill-mask / continued pretraining on legal-domain Uzbek (Latin) text.
• Legal-NLP research: classification by OKOZ code or document type, citation/cross-reference extraction.

Out-of-scope uses

• Any use as authoritative legal text or as a substitute for legal advice. Verify against lex.uz.
• Determining the current in-force status of a provision — this is a pinned snapshot.
• Building systems that present generated output as official legal guidance without a qualified human in the loop.

---

Related datasets

• sukhrobnurali/uzbek-legal-corpus-v1 — the curated trilingual (uz-Latin / uz-Cyrillic / Russian) showcase of the 20 codes + Constitution, with an aligned uz/ru parallel slice. Depth; this corpus is breadth.

---

Source: lex.uz — National Database of Legislation of Uzbekistan. Contact: sukhrobnurali@gmail.com.