README.md

metadata

license: cc-by-4.0
language:
  - kk
  - ru
  - en
multilinguality: multilingual
pretty_name: >-
  AdiletCodex: Trilingual (KK/RU/EN) Article-Level Corpus of Kazakh Legal Codes
  and Laws
tags:
  - legal
  - law
  - legal-nlp
  - kazakhstan
  - kz
  - kazakh
  - russian
  - english
  - legislation
  - statute
  - codes
  - laws
  - adilet
  - article-level
  - trilingual
  - multilingual
  - low-resource
  - central-asia
  - legal-information-retrieval
task_categories:
  - text-retrieval
  - text-classification
  - text-generation
size_categories:
  - 10K<n<100K

AdiletCodex: A Trilingual (Kazakh-Russian-English) Article-Level Corpus of the Codes and Laws of Kazakhstan

A full-text, article-level, trilingual (Kazakh + Russian + English) corpus of the currently in-force codes and substantive laws of the Republic of Kazakhstan, parsed from the official legal database adilet.zan.kz.

• 65,807 article records across 343 documents (24 codes + 319 laws, incl. the Constitution)
• 296 documents present in all three languages (RU 22,112 / KK 22,098 / EN 21,597)
• Adoption years 1991-2026
• Each record: full article text + structural hierarchy (part/section/chapter/article) + numbered points + amendment notes + cross-reference links + rich document metadata (act form, legal sphere, adopting body, registration numbers, dates, status)

To our knowledge this is the first article-structured, trilingual, full-text corpus of Kazakh legislation.

Why this corpus

Kazakh legal NLP has lacked a structured, full-text foundation: existing resources are either document-level metadata, Russian-only, or downstream QA sets built on top of statutes they do not themselves release. AdiletCodex fills that gap with the article as the unit of record, the three official language versions side by side, and the structural + amendment + metadata context needed for legal information retrieval, cross-lingual and RAG grounding, translation-fidelity study, and temporal ("law-as-of-date") modeling.

Related datasets and how AdiletCodex differs

• KazakhLawCorpus (Tleubayeva, Data MDPI 2026; 215,889 records) - document-level metadata only, Kazakh only, no article text. AdiletCodex is complementary and strictly more granular: full text, article-level, trilingual.
• justicedao/ipfs_kazakhstan_laws_ir (HF, ~399k rows) - article-level but Russian only, undocumented, no hierarchy / amendments / status / KK / EN. AdiletCodex adds the two other official languages, the structural hierarchy, amendment notes, status, metadata, and docs.
• olzhasAl/adilet-legal-qa-kz (HF, 13,159 QA pairs, RU+KK) and other Kazakh legal-QA / LoRA sets - downstream tasks, not a source corpus. AdiletCodex provides the underlying statute text those tasks are grounded in.
• Uzbek legal corpus (Tomaris AI) - a methodological twin for a neighbouring jurisdiction (Uzbek-Latin only, no descriptor paper); AdiletCodex is the Kazakh, trilingual counterpart.
• RusLawOD (HF irlspbru/RusLawOD, 304k texts) - Russian national statutory corpus; a language / scope comparator, not Kazakh.

Files

• adiletcodex.jsonl - one article per line (keeps nested paragraphs/notes/links)
• adiletcodex.csv - flat columns

Load

from datasets import load_dataset
ds = load_dataset("json", data_files="adiletcodex.jsonl", split="train")

Schema

doc_id, doc_type, slug, lang, title, status, act_form, act_form_ru, legal_sphere, legal_sphere_ru, adopting_organ, adopting_organ_ru, database_section, database_section_ru, state_registry_number, npa_registration_number, adoption_date, change_date, publication_date, publication_reference, publication_reference_ru, adoption_place, region_action, citation, doc_date, doc_number, part, section, chapter, article_id, article_no, article_title, article_text, paragraphs, notes, links, url

(37 fields total; paragraphs, notes, links are nested and JSONL-only, so the CSV carries the 34 scalar fields. citation is the human-readable source citation, with doc_date / doc_number parsed from it.)

Language of metadata fields. Only title, status, article_title, article_text are truly per-language and are the official adilet translations. The document-"passport" fields come from adilet's Russian-only document card; in v2.0 several are localized to the row's language, each with an *_ru companion preserving the Russian source (Russian rows keep the Russian value):

• Curated term map (hand-reviewed, not machine translation): act_form, adopting_organ, database_section.
• Machine-translated (local Qwen3.5-122B-A10B, thinking disabled) reference metadata - convenience only, not authoritative, with the Russian source kept in the *_ru companion: legal_sphere, publication_reference. The Kazakh output uses correct Kazakh legal terminology (the earlier Qwen2.5-32B Kazakh was replaced). A number-preservation check protects all digits/dates/article numbers in publication_reference (Russian original kept on any mismatch).
• Still Russian-by-source: region_action, adoption_place.

Use status for validity/in-force state. Reproducibility artifacts (term + translation maps, QA-failure log, scripts) are in localization/.

Changed in v2.0 (major): act_form, adopting_organ, database_section are localized per lang via a curated map, and legal_sphere, publication_reference via machine translation (all with *_ru provenance companions); the redundant/mislabeled legal_force field was removed (use status for in-force state). Separately, 18 parser-artifact records (split-off entry-into-force / amendment fragments, a quoted old-code article, truncated compound article numbers, and stale same-language duplicates) were repaired without dropping any wording, so the record count is 65,807 (was 65,825 in v1.0) with 0 duplicate (doc_id, lang, article_no) keys.

License & source

Compilation: CC-BY-4.0. Underlying texts are official public acts of the Republic of Kazakhstan (source: adilet.zan.kz). Unofficial research copy; verify in-force text on adilet.

Citation

Mukhsimbayev, B., Pak, A., & Kuralbayev, A. (2026). AdiletCodex: A Trilingual (Kazakh-Russian-English) Article-Level Corpus of the Codes and Laws of Kazakhstan (v2.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22812625 (concept DOI, always resolves to the latest version). Latest is v2.0 (version DOI 10.5281/zenodo.23260689, published 2026-10-09); v1.0 version DOI: 10.5281/zenodo.22812626.