Uzbek-lang-LM: Unified Foundation & Instruction Corpus

Uzbek-lang-LM is an engineered, linguistically normalized, and decontaminated Uzbek corpus designed for pre-training, continued pre-training, and supervised fine-tuning (SFT) of large language models.

It provides a single, cohesive, modern literary Uzbek language resource adhering strictly to the official Latin script, standardized typography, and logical semantic taxonomies.

---

1. Key Dataset Metrics

Split	Format	Number of Rows	Size (Parquet)	Size (JSONL)
Train	Parquet & JSONL	110,030	170.8 MB	354.8 MB
Validation	Parquet & JSONL	8,619	13.9 MB	29.0 MB
Test (Decontaminated)	Parquet & JSONL	3,094	1.5 MB	4.1 MB
Total Corpus		121,743	186.2 MB	387.9 MB

• Language: Uzbek (uzn_Latn)
• Script: Modern Uzbek Latin (O'zbek lotin alifbosi) with standardized ASCII apostrophe (') for o', g', and the tutuq belgisi.
• License: Apache 2.0 (Permissive for research, academic, and commercial model development).

---

2. Structural Taxonomy & Schema

Every record in the corpus follows a validated 12-attribute unified schema:

{
  "id": "UZLM-TR-0000001",
  "type": "instruction",
  "instruction": "O'rtacha, median va moda tushunchalarini tushuntiring.",
  "input": "",
  "output": "O'rtacha, median va moda – bu statistik ma'lumotlar to'plamidagi markaziy tendensiyaning uchta asosiy ko'rsatkichidir...",
  "messages": [
    {"role": "user", "content": "O'rtacha, median va moda tushunchalarini tushuntiring."},
    {"role": "assistant", "content": "O'rtacha, median va moda – bu statistik ma'lumotlar to'plamidagi..."}
  ],
  "text": "User: O'rtacha, median va moda tushunchalarini tushuntiring.\n\nAssistant: O'rtacha, median va moda...",
  "topic": "general_instruction",
  "source": "uzbek_instruction_archive",
  "license": "apache-2.0",
  "is_synthetic": false,
  "quality_score": 1.0,
  "split": "train"
}

Identifier Protocol

Identifiers are deterministically generated according to split and ordinal sequence:

• UZLM-TR-XXXXXXX for Training records (UZLM-TR-0000001 to UZLM-TR-0110030)
• UZLM-VA-XXXXXXX for Validation records (UZLM-VA-0000001 to UZLM-VA-0008619)
• UZLM-TE-XXXXXXX for Decontaminated Test records (UZLM-TE-0000001 to UZLM-TE-0003094)

---

3. Core Capabilities & Domains

The corpus encompasses 14 balanced domains:

1. General Knowledge: World geography, physical phenomena, and civilizations.
2. History: Central Asian dynasties, historical figures (Amir Timur, Ulugh Beg, Babur, Khwarizmi, Ibn Sina, Biruni), and global milestones.
3. Uzbek Literature & Culture: Alisher Navoiy's works, Jadid authors (Qodiriy, Cho'lpon, Fitrat), 20th-century classics, folk epics, and applied arts.
4. Natural Sciences: Physics, chemistry, genetics, cell biology, and astronomy.
5. Computer Science & Coding: Python, JavaScript, and SQL programming with algorithmic analysis and native Uzbek explanations.
6. Jurisprudence & Law: Constitutional law, civil contracts, labor regulations, and civic rights.
7. Medicine & Health: Anatomy, preventive medicine, cardiovascular health, nutrition, and emergency first aid.
8. Business & Management: Startup lifecycles (MVP), financial cash flow, digital marketing, and strategic risk management.
9. Daily Communication: Professional etiquette, job interview dialogues, customer relations, and formal correspondence.
10. Translation & Linguistics: Bidirectional Uzbek-English and Uzbek-Russian translation with contextual linguistic notes.
11. Mathematical Reasoning: Step-by-step arithmetic, geometry, algebra, work productivity, and probability with verified calculations.
12. Summarization: Dense informative passages with structured executive summaries.
13. Ethical AI Assistance: Polite, principled, constructive responses adhering to safety standards.
14. Linguistic Benchmarks: Isolated orthographic and grammatical evaluation sets.

---

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, ts at word beginnings -> s, ў -> o', ғ -> g', қ -> q, ҳ -> h).
• Apostrophe Unification: Normalized all unicode variant quotes and modifiers (ʻ, ’, ‘, `, ´, ʼ, ′) to the standard ASCII apostrophe (') for o', g', and the tutuq belgisi (ma'no, sa'y, e'tibor).
• Encoding & Text Cleansing: Enforced Unicode NFC normalization, resolved UTF-8 mojibake, stripped HTML tags, and removed control characters.

Multi-Stage Quality Filtering

• Language Verification: Verified using OpenLID fastText (facebook/fasttext-language-identification) with confidence threshold p >= 0.90 for Uzbek (__label__uzn_Latn).
• Safety & PII Screening: Filtered personal information (emails, phone numbers, identity numbers), toxicity, and harmful generation requests.
• Mathematical Validation: Programmatically checked every arithmetic equation to guarantee calculation accuracy.

Deduplication

• Exact Hash Deduplication: SHA-256 canonical text hashing.
• Near-Deduplication: MinHash LSH (64 permutations, Jaccard similarity threshold >= 0.85).

Strict Decontamination

The evaluation and test splits (test.parquet) are fully isolated from the training corpus, guaranteeing zero contamination 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-lang-lm")

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

# Access training subset for instruction fine-tuning
train_ds = dataset["train"]
print(f"Total training examples: {len(train_ds)}")

---

6. Dataset Structure

.
├── data/
│   ├── train.parquet        # 110,030 examples (170.8 MB)
│   ├── train.jsonl          # 110,030 examples (354.8 MB)
│   ├── validation.parquet   # 8,619 examples (13.9 MB)
│   ├── validation.jsonl     # 8,619 examples (29.0 MB)
│   ├── test.parquet         # 3,094 examples (1.5 MB)
│   └── test.jsonl           # 3,094 examples (4.1 MB)
├── .gitattributes
└── README.md

---

7. Citation

@dataset{uzbek_lang_lm_2026,
  title={Uzbek-lang-LM: Unified Foundation & Instruction Corpus for Uzbek Language Models},
  author={Uzbek NLP Initiative},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/ulugbek2013/uzbek-lang-lm}}
}