ansarzeinulla/Nogai-Russian-SFT-Biblical-v1

📚 Nogai-Russian SFT Biblical Corpus v1

Curing Catastrophic Forgetting in Zero-Resource Turkic LLM Adaptation

A structurally flawless, human-verified instruction-tuning dataset engineered to recover latent prompt-adherence following unstructured Continuous Pre-Training (CPT).

🔬 Overview & The "Catastrophic Forgetting" Bottleneck

When adapting frontier Large Language Models (e.g., Qwen, Llama) to entirely new, isolated languages via Continuous Pre-Training (CPT) on raw text, the architectures mathematically suffer from Catastrophic Forgetting. While the gradient optimizer successfully learns the local Turkic morphological probability distributions, it overwrites the model's instruction-following manifold. The model loses its ability to obey system prompts, format outputs, or act as a conversational agent.

The Nogai-Russian SFT Biblical Corpus v1 was explicitly engineered for Phase 2 Supervised Fine-Tuning (SFT) to cure this collapse. It perfectly maps the newly acquired Nogai vocabulary back onto the native ChatML instruction vectors, establishing a rigorous mathematical baseline for cross-lingual semantic alignment.

---

📊 Dataset Specifications

Metric	Specification
Total File Size	6.49 MB
Total Rows (Sequences)	4,310
Data Format	ChatML formatted .jsonl
Languages	Nogai (nog), Russian (ru)
Splits	train (81%), validation (19%)

---

⚙️ Data Engineering & Syntactic Preservation

Constructing parallel data for zero-resource languages frequently fails due to Grammatical Boundary Mismatch. Slavic and Turkic languages lack 1:1 syntactic mapping. Standard extraction scripts that truncate sequences by token count slice sentences into ungrammatical fragments, destroying cross-attention mappings and causing associative hallucination.

To bypass this, we implemented a highly rigorous extraction pipeline:

1. Human-Verified Gold Standard: Exclusively sourced from human-translated religious archives (Institute for Bible Translation). This guarantees absolute semantic equivalence and zero machine-translation (MT) contamination.
2. Proportional Sentence-Boundary Chunking: We developed a custom algorithmic chunker that isolates syntactically complete sentences using terminal delimiters. It proportionally groups sequences to ensure every training sample contains grammatically intact thought blocks, completely eliminating un-severed cross-lingual vectors.
3. Bidirectional Tensor Alignment: To symmetrically stabilize the LLM's multi-head attention mechanisms, the SFT prompts are mathematically mirrored. Exactly 50% of the instruction prompts request Russian $\rightarrow$ Nogai, while the remaining 50% request Nogai $\rightarrow$ Russian.

---

📝 Schema & ChatML Structure

The data is pre-formatted into the strict ChatML dictionary schema natively required by modern instruct-tuned models (e.g., Qwen-2.5, Llama-3).

{
  "messages": [
    {"role": "system", "content": "You are a highly accurate bilingual translator for Russian and Nogai."},
    {"role": "user", "content": "Переведи этот текст на ногайский язык: В начале сотворил Бог небо и землю."},
    {"role": "assistant", "content": "Енъ баслап Кудай коькти эм ерди яратты."}
  ]
}

---

⚠️ Scientific Limitations: Structural-Semantic Overfitting

We transparently define the "Data Ceiling" of this dataset. Because this is a micro-dataset ($N \approx 4,310$ chunked sequences) derived exclusively from theological/narrative domains, models fine-tuned solely on this dataset will exhibit Structural-Semantic Overfitting.

What this means in practice:

• The model will structurally understand how to execute prompts flawlessly (curing the ChatML formatting collapse).
• However, when prompted with open-domain, out-of-distribution (OOD) modern conversational text (e.g., medical or technological queries), the predictive routing will map to the nearest high-probability cluster in the SFT weights, defaulting to biblical lexicon (e.g., repeating archaic vocabulary or theological concepts).

Note: Achieving true open-domain zero-shot fluency requires leveraging this baseline to bootstrap synthetic back-translation at a scale of 10k+ sequences.

---

🚀 Quick Start: Loading the Dataset

Ready for plug-and-play into PyTorch, MLX, or Hugging Face TRL (Transformer Reinforcement Learning) pipelines:

from datasets import load_dataset

# Load the SFT ChatML dataset
dataset = load_dataset("ansarzeinulla/Nogai-Russian-SFT-Biblical-v1")

# Preview the structure
train_data = dataset["train"]
print(f"Loaded {len(train_data)} grammatically intact ChatML chunks.")
print(train_data[0]["messages"])

---

📚 Citation

If you utilize our SFT methodology, boundary chunking algorithms, or this dataset, please cite our ACM TALLIP / arXiv paper:

@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_russian_sft_2026,
  author       = {Zeinulla, Ansar},
  title        = {Nogai-Russian SFT Biblical Corpus v1: Curing Catastrophic Forgetting in Zero-Resource LLM Adaptation},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ansarzeinulla/Nogai-Russian-SFT-Biblical-v1}},
}

Curated and maintained by @ansarzeinulla • Nazarbayev University