🧠 Pashto-Reasoning-12K

The first large-scale Chain-of-Thought (CoT) reasoning dataset for the Pashto language.

Pashto is spoken by over 60 million people worldwide, yet it remains critically underrepresented in modern NLP and reasoning research. This dataset is a pioneering effort by iPashto.ai to bring deep reasoning and logical "thinking" capabilities to Pashto Language Models.

---

📋 Dataset Summary

Field	Value
Language	Pashto (ps)
Total Rows	13,317
Data Format	JSONL / Parquet
Core Focus	Chain-of-Thought (CoT) & Reasoning
License	Apache 2.0

---

📐 Dataset Structure

Each sample follows a modern reasoning architecture with 4 distinct fields:

• system: Fixed persona instruction for the Pashto assistant.
• user: Input query, task, or question in Pashto (or mixed).
• thought: The internal reasoning chain or "step-by-step logic" (CoT).
• assistant: The final clear and concise answer in Pashto.

JSON Example

{
  "system": "You are a helpful Pashto assistant...",
  "user": "د «ریاضیي» اصطلاح څه معنی لري؟",
  "thought": "User asks about the definition of 'mathematical'. I will explain its etymology and common usage in Pashto.",
  "assistant": "«ریاضیي» د ریاضي پوه، یا د ریاضي اړوند زده کړې لپاره کارول کیږي."
}

---

🚀 Usage & Integration

Load via Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset("nassimjp/Pashto-Reasoning-12K")
print(ds["train"][0])

Fine-tuning with ChatML Template

This dataset is ideal for training reasoning models (like DeepSeek-R1 style or Qwen) using the following template:

<|system|>{{system}}<|user|>{{user}}<|thought|>{{thought}}<|assistant|>{{assistant}}

---

⚠️ Known Limitations & Curation

• Mixed-Language Thought: Many reasoning traces (thought field) contain English logic to solve Pashto problems. This is designed to leverage the cross-lingual reasoning capabilities of base LLMs.
• Task Diversity: Covers Grammar, Morphology, Lexical Semantics, Translation, and Cultural Knowledge.
• Version: This is v1.0. Some shallow reasoning patterns may exist and will be improved in future releases.

---

🌍 Why Pashto?

Despite being a major Indo-Iranian language, Pashto lacks high-quality instruction-tuning and reasoning data. iPashto.ai, based in Saitama, Japan, is dedicated to:

1. Building native Pashto datasets.
2. Developing open-source models (Ghanam, Roshan, Baran).
3. Bridging the gap between low-resource languages and SOTA AI.

---

📜 Citation

If you use this dataset in your research, please cite:

@dataset{nassimjp2026pashto12k,
  author    = {Nassim Nasibullah},
  title     = {Pashto-Reasoning-12K: A Chain-of-Thought Dataset for Pashto},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {[https://huggingface.co/datasets/nassimjp/Pashto-Reasoning-12K](https://huggingface.co/datasets/nassimjp/Pashto-Reasoning-12K)}
}

---

🤝 Contact & Contributions

Created and maintained by @nassimjp. Contributions, feedback, and collaborations are welcome via the community tab! ```