Uzbek-law-LM: Authoritative Legal Corpus & SFT Benchmark

Uzbek-law-LM is an engineered, legally grounded, and decontaminated Uzbek legal corpus designed for pre-training, continued pre-training, and supervised fine-tuning (SFT) of large language models serving the Republic of Uzbekistan's legislative, judicial, executive, and civic sectors.

It provides a single, cohesive, authoritative legal resource adhering strictly to the official Uzbek Latin script, standardized canonical typography, and structured statutory citation provenance.

---

1. Key Dataset Metrics

Split	Format	Number of Rows	Size (Parquet)	Size (JSONL)
Train	Parquet & JSONL	65,035	145.7 MB	479.6 MB
Validation	Parquet & JSONL	5,261	12.1 MB	39.7 MB
Test (Decontaminated)	Parquet & JSONL	6,586	12.9 MB	44.0 MB
Total Corpus		76,882	170.7 MB	563.4 MB

• Language: Uzbek (uzn_Latn) with official parallel constitutional provisions in Russian and English.
• Script: Canonical Uzbek Latin (Oʻzbek lotin alifbosi) with standardized modifier letter turned comma (ʻ U+02BB) for oʻ and gʻ, and modifier letter apostrophe (ʼ U+02BC) for the tutuq belgisi (maʼno, saʼy, eʼtibor).
• License: Apache 2.0 (Permissive for research, governance, academic, and commercial model development).

---

2. Structural Taxonomy & Schema

Every record in the corpus follows a validated 23-attribute unified schema:

{
  "id": "UZLAW-TR-0048461",
  "type": "instruction",
  "instruction": "Qonunchilikka koʻra, yer uchastkalarini auksion orqali ajratish tartibi qanday?",
  "input": "",
  "output": "**Oʻzbekiston Respublikasining Yer kodeksi**ga muvofiq, yer uchastkalari tadbirkorlik faoliyati uchun faqat elektron onlayn-auksion vositasida ajratiladi...",
  "messages": "",
  "text": "Savol:\nQonunchilikka koʻra, yer uchastkalarini auksion orqali ajratish tartibi qanday?\n\nJavob:\n**Oʻzbekiston Respublikasining Yer kodeksi**ga muvofiq...",
  "query": "",
  "legal_domain": "land",
  "audience": "citizen",
  "citation_document_type": "Kodeks",
  "citation_issuing_body": "Oʻzbekiston Respublikasi Oliy Majlisi",
  "citation_document_number": "Yer kodeksi",
  "citation_adoption_date": "1998-04-30",
  "citation_article": "17-modda",
  "citation_source_url": "https://lex.uz/docs/152653",
  "citation_access_date": "2026-09-19",
  "legal_status": "active",
  "source": "Uzbek-law-LM-Grounding",
  "license": "apache-2.0",
  "is_synthetic": true,
  "quality_score": 0.98,
  "split": "train"
}

Identifier Protocol

Identifiers are deterministically generated according to split and ordinal sequence:

• UZLAW-TR-XXXXXXX for Training records (UZLAW-TR-0000001 to UZLAW-TR-0065035)
• UZLAW-VA-XXXXXXX for Validation records (UZLAW-VA-0000001 to UZLAW-VA-0005261)
• UZLAW-TE-XXXXXXX for Decontaminated Test records (UZLAW-TE-0000001 to UZLAW-TE-0006586)

---

3. Core Legal Capabilities & Domains

The corpus encompasses 56,382 authoritative pre-training statutory provisions and 20,500 verified supervised fine-tuning instruction pairs across 11 specialized categories:

1. Legal Q&A (2,500 pairs): Factual questions on statutory requirements answered with article citations.
2. Legal Terminology (1,800 pairs): Dual explanations for citizens (plain language) and jurists (technical doctrine).
3. Summarization (2,000 pairs): Executive summaries (boshqaruv xulosasi) for ministers and parliamentary deputies.
4. Comparative Law (1,500 pairs): Diachronic legal changes (1992 vs 2023 Constitution, 1995 vs 2023 Labor Code).
5. Document Drafting (2,200 pairs): Petitions (daʼvo ariza), complaints (shikoyat), contracts, and deputy requests.
6. Legal Reasoning (IRAC) (2,200 pairs): Structured dispute resolution (Muammo → Qoida → Tahlil → Xulosa).
7. Legislative Work (1,800 pairs): Regulatory impact analysis (TSTB) and anti-corruption screening.
8. Public Services (2,000 pairs): Citizen administrative workflows grounded in my.gov.uz regulations.
9. Classification & NER (2,000 pairs): Legal domain identification and statutory entity extraction.
10. Trilingual Translation (1,500 pairs): Grounded UZ ↔ EN and UZ ↔ RU parallel constitutional articles.
11. Safe Legal Refusals (1,000 pairs): Principled refusals of illegal schemes citing statutory prohibitions.

Legal Branches Covered

• Constitutional Law: 2023 amended Constitution (155 articles) and electoral legislation.
• Civil & Business Law: Property, contracts, torts, and corporate governance.
• Criminal & Administrative Law: Criminal Code, Administrative Responsibility Code.
• Labor Law: 2023 Labor Code (599 articles) and social security regulations.
• Tax & Customs: Tax Code, customs clearance, and fiscal exemptions.
• Land & Family: Urban planning, land cadastre, marriage, and inheritance.
• Procedural Law: Civil, Criminal, Economic, and Administrative Court procedures (FPK, JPK, IPK, MSYP).

---

4. Engineering & Curation Standards

Linguistic Normalization

• Script Standardization: Context-aware Cyrillic-to-Latin transliteration implemented with phonetic accuracy (e.g. е after vowels/signs → ye, ц → ts/s, ў → oʻ, ғ → gʻ, қ → q, ҳ → h).
• Canonical Typography: Strict standardization of oʻ and gʻ using modifier letter turned comma (ʻ U+02BB), and tutuq belgisi using modifier letter apostrophe (ʼ U+02BC). All variant quotes and modifiers normalized.
• Encoding & Text Cleansing: Enforced Unicode NFC normalization, resolved UTF-8 mojibake, stripped HTML tags, removed control characters, and eliminated non-breaking spaces.

Multi-Stage Quality Filtering

• Language Verification: Uzbek script heuristics with confidence thresholds ensuring >= 90% Uzbek Latin content per document.
• PII Anonymization: Automated regex engine masking passport numbers ([PASSPORT_RAQAMI]), 14-digit PINFL/JSHSHIR codes ([JSHSHIR_RAQAMI]), credit card numbers, and phone numbers.
• Terminology Enforcement: Complete eradication of colloquial Russianisms (iskatel → daʼvogar, otvetchik → javobgar, podzakonniy akt → qonunosti hujjati) via a 25-term authoritative glossary.

Deduplication

• Exact Hash Deduplication: xxHash64 canonical text hashing.
• Near-Deduplication: MinHash LSH (64 permutations, banded buckets, Jaccard similarity threshold >= 0.85).

Strict Decontamination

Partitioned strictly by legal act and document identity. All provisions from a given statute are confined to a single split, guaranteeing zero data leakage during model benchmarking.

---

5. Usage with the datasets Library

from datasets import load_dataset

# Load all splits (train, validation, test)
dataset = load_dataset("ulugbek2013/uzbek-law-lm")

# Inspect sample record
print(dataset["train"][0])

# Access training subset for instruction fine-tuning
sft_train = dataset["train"].filter(lambda x: x["type"] != "pretrain_text")
print(f"Total SFT training examples: {len(sft_train)}")

---

6. Dataset Structure

.
├── data/
│   ├── train.parquet        # 65,035 examples (145.7 MB)
│   ├── train.jsonl          # 65,035 examples (479.6 MB)
│   ├── validation.parquet   # 5,261 examples (12.1 MB)
│   ├── validation.jsonl     # 5,261 examples (39.7 MB)
│   ├── test.parquet         # 6,586 examples (12.9 MB)
│   └── test.jsonl           # 6,586 examples (44.0 MB)
├── .gitattributes
└── README.md

---

7. Citation

@dataset{uzbek_law_lm_2026,
  title={Uzbek-law-LM: Authoritative Uzbek Legal Corpus \& SFT Benchmark},
  author={Uzbek NLP \& Legal Intelligence Initiative},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/ulugbek2013/uzbek-law-lm}}
}