ansarzeinulla/Nogai-Unified-Corpus-v1

📚 Nogai Unified Corpus (NUC) v1

The First Mathematically Isolated, Multi-Domain Corpus for the Endangered Nogai Language

Engineered to break the data scarcity bottleneck in zero-resource Turkic NLP, enabling high-fidelity Continuous Pre-Training (CPT) without cross-lingual contamination.

🔬 Overview & The Data Contamination Problem

Historically, the critically endangered Nogai language (a Turkic language spoken primarily in the North Caucasus region) has been classified as a "zero-resource" language. Adapting frontier LLMs to Nogai via web-scraping inevitably captures high-resource linguistic contamination (predominantly Russian). This contamination distorts token frequency distributions, causing self-attention mechanisms in models like Qwen or Llama to default to Slavic morphological rules during generation.

The Nogai Unified Corpus v1 (NUC) was engineered specifically to solve this. It is the largest publicly available, meticulously curated, and mathematically isolated textual dataset for the Nogai language, designed specifically to inject structural Cyrillic Turkic morphology into modern NLP architectures.

---

📊 Dataset Specifications

Metric	Specification
Total File Size	35.6 MB
Total Rows (Sequences)	163,531
Estimated Tokens	~4,000,000
Data Format	.jsonl
Splits	train (95%), validation (5%)

---

🏗️ Data Provenance & Domain Anchors

To prevent Domain Collapse during Continuous Pre-Training, the corpus was synthesized across three mathematically balanced semantic anchors:

1. Journalistic & Formal Anchor (The News Archives)

Constitutes the structural majority of the corpus. Extracted via custom, highly optimized PDF parsing pipelines from over 600 digitized digital issues spanning a decade of publication (2015–2026).

• "Голос степи" / "Шоьл тавысы" (Voice of the Steppe): 500+ issues published in Dagestan, Russia.
• "Ногай давысы" (Nogai Voice): 100+ issues published in Karachay-Cherkessia, Russia.

2. Narrative & Semantic Anchor (IBT Archives)

To accurately map fundamental human semantic concepts (e.g., life, emotions, nature, warnings) with high structural integrity, human-translated religious and narrative archives were utilized.

• Source: Institute for Bible Translation (IBT) Russia.

3. Factual & Encyclopedic Anchor (Wikimedia)

Injected to ground the model's self-attention layers in objective, encyclopedic syntax, modern definitions, and geographical formatting.

• Source: Wikimedia Incubator (Wp/nog).

---

⚙️ Preprocessing & Mathematical Isolation

Raw data extraction is insufficient for zero-resource languages. The NUC v1 underwent a highly rigorous normalization pipeline:

1. OCR Artifact Eradication: Heuristic cleaning of broken Cyrillic encodings.
2. Cross-Lingual Filtering: Standard Russian stop-word frequency filtering was applied to systematically purge high-resource contamination and isolate the true Nogai Cyrillic vocabulary.
3. Boundary Segmentation: Full sentence/paragraph segmentation.
4. Batch Diversification: The entire dataset was heavily randomized/shuffled to ensure maximum variance during gradient descent batching.

---

🚀 Quick Start: Loading the Dataset

The dataset is ready for immediate deployment in PyTorch/JAX/MLX training loops via the Hugging Face datasets library.

from datasets import load_dataset

# Load the dataset (automatically downloads and caches)
dataset = load_dataset("ansarzeinulla/Nogai-Unified-Corpus-v1")

# Access the training split
train_data = dataset["train"]
valid_data = dataset["validation"]

print(f"Total training sequences: {len(train_data)}")
print(f"Sample sequence: {train_data[0]['text']}")

---

🎯 Primary Use Cases

• Parameter-Efficient Continuous Pre-Training (CPT): Expanding the vocabularies of massive frontier models (e.g., Qwen-2.5, Llama-3.2) via QLoRA/LoRA native training pipelines.
• Zero-Shot NLP Benchmarking: Establishing perplexity baselines and Natural Language Understanding (NLU) evaluation sets for endangered Turkic languages.
• Morphological Analysis: Training sequence-to-sequence NMT models for agglutinative suffix mapping.

---

📚 Citation

If this dataset accelerates your research in zero-resource adaptation or Turkic NLP, please cite the parent ACM paper and the dataset repository:

@article{zeinulla2026nogaillm,
  title={NogaiLLM: Parameter-Efficient Continuous Pre-Training and Architectures of Catastrophic Forgetting in Zero-Resource Turkic Languages},
  author={Zeinulla, Ansar},
  journal={arXiv preprint arXiv:2607.xxxxx},
  year={2026},
  publisher={Nazarbayev University / ACM TALLIP}
}

@dataset{nogai_unified_corpus_2026,
  author       = {Zeinulla, Ansar},
  title        = {Nogai-Unified-Corpus-v1: A Comprehensive Dataset for Low-Resource Turkic Language Modeling},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ansarzeinulla/Nogai-Unified-Corpus-v1}},
}

Curated and maintained by @ansarzeinulla • Nazarbayev University