This is a Pashto translation of CoT instructions of OpenOrca dataset.

Pashto_OrcaCoT: A High-Quality Chain-of-Thought Dataset for Pashto Reasoning

Pashto_OrcaCoT is the premier open-source, high-quality Chain-of-Thought (CoT) dataset natively optimized for the Pashto language. Built upon rigorous multi-step logic and structured reasoning patterns inspired by the Orca framework, this dataset bridges the gap between international LLM benchmarks and Pashto-native artificial intelligence alignment.

It serves as the foundational reasoning layer for the iPashto.ai engine and acts as a critical alignment asset for advanced models, including the upcoming Qwen3-1.7B-Pashto-Gold pipeline.

---

🌟 Key Features

• Structured CoT: Features multi-step logic explicit flows, breakdown reasoning, and rigorous, verifiable formatting.
• Pashto-Native Alignment: System prompts and instructions are perfectly contextualized to respect Pashto linguistic nuances, cultural subtexts, and stylistic accuracy.
• High-Quality Integrity: Optimized to counteract hallucinations, maximize logical consistency, and ensure deep reasoning depth across mathematics, logic, and common-sense QA.
• Zero-Scraping Strategy: Curated with precision, bypassing noisy web-scraping to deliver gold-standard, production-ready instruction tuning.

---

📊 Dataset Schema

Each record follows a strict three-field sequence designed for seamless Integration during model training:

Field	Type	Description
instruction	string	The system prompt or agent persona defining the behavioral boundaries.
input	string	The user query, task statement, or reasoning puzzle in Pashto.
output	string	The gold-standard response containing the step-by-step reasoning logic and final answer.

Data Example

{
  "instruction": "تاسو د مصنوعي ځیرکتیا (AI) یو مرستیال یاست چې له خلکو سره د معلوماتو په موندلو کې مرسته کوي. کاروونکی له تاسو څخه پوښتنه کوي. ستاسو دنده دا ده چې تر خپله وسه په پوره صداقت ځواب ورکړئ. د ځواب ورکولو پر مهال ګام په ګام فکر وکړئ او د خپل ځواب لپاره منطقي دلایل وړاندې کړئ.",
  "input": "جیمز د 5 اصلي کرکټرونو او 4 ملاتړي کرکټرونو سره د تلویزیون خپرونه پرمخ وړي. هغه په ​​هر قسط کې د ملاتړي کرکټرونو $ 15,000 تادیه کوي. هغه اصلي کرکټرونه درې چنده ډیر ورکړي. هغه په ​​هره برخه کې څومره پیسې ورکوي؟ راځئ چې د امکان تر حده مشخص وي.",
  "output": "جیمز د ملاتړ کونکي کرکټرونو په هره برخه کې $15,000 تادیه کوي. له هغه ځایه چې 4 ملاتړي کرکټرونه شتون لري، دا دوی ته په ټوله برخه کې 4 *$ 15,000 = $60,000 ورکوي.\n\nاصلي کرکټرونه درې ځله ډیر تادیه کیږي. نو، هر اصلي کرکټر په هره برخه کې$ 3 * $15,000 =$ 45,000 ترلاسه کوي.\n\nدلته 5 اصلي کرکټرونه دي، نو هغه دوی ته په هر قسط کې 5 * $45,000 =$ 225,000 تادیه کوي.\n\nپه مجموع کې، جیمز ته $225,000 (اصلي کرکټرونه) +$ 60,000 (ملاتړي کرکټرونه) = $ 285,000 په هره قسط ورکول کیږي."
}

---

🛠️ Motivation & Vision

Historically, low-resource languages like Pashto have suffered from noisy, poorly translated datasets that degrade an LLM's capability to reason coherently. Pashto_OrcaCoT marks the end of the zero-scraping era for Pashto AI, providing a clean, hand-crafted, and machine-optimized environment for supervised fine-tuning (SFT). The goal is to allow Pashto language models to perform complex chain-of-thought deductions at par with English and other globally rich resource environments.

---

📜 Citation & Attributions

If you use this dataset in your research, model fine-tuning pipelines, or evaluation benchmarks, please cite it using the following format:

@dataset{nassimjp_pashto_orcacot_2026,
  author       = {Nassim},
  title        = {Pashto_OrcaCoT: A High-Quality Chain-of-Thought Dataset for Pashto Reasoning},
  year         = {2026},
  publisher    = {Hugging Face},
  journal      = {Hugging Face Datasets},
  howpublished = {\url{[https://huggingface.co/datasets/nassimjp/Pashto_OrcaCoT](https://huggingface.co/datasets/nassimjp/Pashto_OrcaCoT)}}
}

---

🤝 Feedback and Collaboration

For issues, contributions, or collaborations regarding Pashto reasoning tools and native LLM development, please open an issue in the repository or contact us via iPashto.ai.