nassimjp/Pashto-Free-Hand-Reasoning-Dataset

Pashto Free-Hand Reasoning SFT Dataset 🧠♻️

This dataset contains high-quality, long-form SFT (Supervised Fine-Tuning) conversational data in Pashto, featuring unconstrained, natural model reasoning (<think> blocks) paired with standardized chat responses.

🔄 The 3R Approach (Recycle, Reuse, Reason)

Instead of discarding legacy QA pairs, this dataset follows a 3R data philosophy:

• Recycle: Taking older, simple, or raw legacy Pashto questions.
• Reuse: Re-processing them through state-of-the-art reasoning engines.
• Reason: Upgrading simple static text into dynamic, step-by-step chain-of-thought (CoT) datasets containing rich <think> blocks.

Dataset Structure

The dataset is formatted in standard multi-turn chat format using JSONL, where each record contains system, user, and assistant roles. The assistant's response incorporates native reasoning paths inside <think> tags followed by the final refined response:

{
  "messages": [
    {"role": "system", "content": "ته یو مرستندویه او پوه پښتو ژبی مرستندویه ماډل یې."},
    {"role": "user", "content": "پوښتنه دلته وي..."},
    {"role": "assistant", "content": "<think>\nطبیعي او ژور فکر او استدلال دلته وي...\n</think>\n\nاصلي او معیاري ځواب دلته وي."}
  ]
}

Data Generation Pipeline

• Source Questions: Extracted from legacy and clean Pashto QA pairs to upgrade and elevate simple data into deep, reasoning-enabled instructions.
• Reasoning Engine: Generated using advanced open-weights models through a model-agnostic local mining pipeline with llama.cpp and unrestricted context limits.
• Format: OpenAI-compatible standard chat ML format.

🙏 Acknowledgments & Thanks to Developers

A massive shoutout and heartfelt thanks to the open-source AI community and developers worldwide! Special appreciation goes to:

• The llama.cpp team for building high-performance, flexible local inference engines that make heavy free-hand reasoning pipelines possible.
• The creators of open-weights models (such as Qwen and fine-tuned variants) for pushing the boundaries of multilingual and low-resource intelligence.
• All open-source contributors and data engineers striving to elevate the Pashto language in the AI ecosystem.

License

This project is open-sourced under the Apache 2.0 License.

Use Cases

• Training and fine-tuning reasoning-capable Large Language Models (LLMs) for the Pashto language.
• Enhancing native low-resource chain-of-thought (CoT) capabilities.