nassimjp/pashto-reasoning-bazaar-v1

Pashto Reasoning Bazaar Dataset (iPashto.ai v5.5)

The world's first comprehensive Pashto dataset specifically designed for Chain of Thought (CoT) reasoning and mental mathematics. This dataset simulates real-world bazaar transactions across 20 provinces of Afghanistan, providing a rich ground for training models in Pashto linguistic nuances and logical deduction.

Dataset Statistics

• Total Samples: 10,000+
• Provinces Covered: 20
• Unique Products: 48 (Local & Imported)
• Reasoning Strategies: 7 unique mental math methods
• Response Styles: 7 natural conversational variations
• Format: ShareGPT/OpenAI compatible (Messages format)

Dataset Structure

The data is structured to be compatible with modern instruction-tuning frameworks (like Axolotl, Llama-Factory, or Swift):

{
  "messages": [
    {
      "role": "system",
      "content": "ته د پښتو ژبې یو هوښیار مرستیال یې، چې نوم دی \"پلوشه\" (Palwasha)..."
    },
    {
      "role": "user",
      "content": "که د کندهار په بازار کې 3 من انار (هر من 620 افغانۍ) واخلئ، ټول څومره پیسې ورکوئ؟"
    },
    {
      "role": "assistant",
      "content": "<thought>\nStep 1: Price of 1 Man of Pomegranates = 620 AFN\nStep 2: 3 × 620 = (3×600) + (3×20) = 1800 + 60 = 1860\nResult: 1860 AFN\n</thought>\n\nتاسو به په ټولیزه توګه 1860 افغانۍ ورکوئ."
    }
  ]
}

Reasoning Strategies

To move beyond simple rote multiplication, the dataset incorporates various mental math strategies to train the model's "thinking" process:

1. Identity: Basic multiplication by 1 ($1 \times n = n$).
2. Direct: Standard multiplication table lookup.
3. Multiply by 10: Appending zeros ($n \times 10$).
4. Multiply by 5: The half-of-ten rule ($(n \times 10) / 2$).
5. Multiply by 9: The compensation rule ($(n \times 10) - n$).
6. Decomposition: Breaking down numbers ($(a+b) \times c = ac + bc$).
7. Compensation: Rounding up and subtracting ($(a+1) \times b - b$).

Geographic Coverage (20 Provinces)

Kandahar, Nangarhar, Paktia, Herat, Kabul, Bamiyan, Kunduz, Helmand, Ghazni, Farah, Sar-e Pol, Badakhshan, Khost, Logar, Jowzjan, Takhar, Faryab, Balkh, Baghlan, Nimruz.

Product Categories

• Local Fruits: Pomegranate, Orange, Apple, Melon, Watermelon, Mango, Grapes, Raisins, Apricot, Mulberry, Almond, Pistachio, Honey, Saffron.
• Local Vegetables: Eggplant, Tomato, Potato, Onion, Pumpkin, Okra, Cauliflower, Chili, Greens, Zucchini, Wheat, Corn, Chickpeas, Beans.
• Imported Goods: Malaysian Oil, Indian Sugar, Iranian Fuel, Pakistani Tea, Turkish Olive Oil, Chinese Rice, Russian Wheat, UAE Oil.

How to Use

1. Load the Dataset

from datasets import load_dataset

dataset = load_dataset("nassimjp/pashto-reasoning-bazaar-v1")
print(dataset["train"][0])

2. Fine-tuning with Qwen/Llama

The dataset is optimized for chat-template application:

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
dataset = dataset.map(lambda x: {"text": tokenizer.apply_chat_template(x["messages"], tokenize=False)})

License

This dataset is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

Summary:

• Attribution: You must give appropriate credit (Nasibullah Nassim/iPashto.ai).
• Non-Commercial: You may not use the material for commercial purposes.
• Share: You are free to copy and redistribute the material in any medium or format.

Citation

@dataset{nassim2026pashto,
  author = {Nasibullah Nassim},
  title = {Pashto Reasoning Bazaar Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/nassimjp/pashto-reasoning-bazaar-v1},
  note = {First-ever complete Pashto Chain-of-Thought reasoning dataset for bazaar transactions.}
}

About the Author

Developed by Nasibullah Nassim, an AI Developer and IT Specialist based in Saitama, Japan. This project is part of the iPashto.ai initiative to build missing AI infrastructure for the Pashto language.

Project Website: iPashto.ai
Contact: Hugging Face Profile

---