nassimjp/ghanam-reasoning-dataset

Ghanam Reasoning Dataset (Pashto-English)

Overview

Ghanam Reasoning Dataset is the first specialized Pashto-English translation dataset designed for Chain-of-Thought (CoT) and reasoning models. It focuses on proverbs, wisdom, and complex linguistic structures, providing a step-by-step thinking process (<thought>) before the final translation.

This dataset is part of the iPashto.ai initiative, supporting models like Ghanam-1.B and Baran.

Dataset Structure

Each entry follows a chat-based format with a system message in English and user interaction in Pashto:

• System: Defines the role as an expert translator with reasoning capabilities.
• User: Input sentence in Pashto (prefixed with "دا جمله په انګلیسي وژباړه:").
• Assistant:
  • <thought>: A multi-step logical analysis of the Pashto grammar, lexical meaning, and cultural context.
  • Final Translation: The accurate English equivalent.

Example Entry

{
  "messages": [
    {
      "role": "system",
      "content": "You are an expert Pashto-English translator. Before providing the translation, think step-by-step about the meaning, cultural context, and grammar of the Pashto sentence."
    },
    {
      "role": "user",
      "content": "دا جمله په انګلیسي وژباړه: د صبر کاسه تل خوږه وي"
    },
    {
      "role": "assistant",
      "content": "<thought>\n1. Lexical Analysis: Identify 'صبر' (patience) and 'کاسه' (bowl/cup).\n2. Context: A proverb meaning patience leads to good results.\n3. Drafting: Mapping to 'The bowl of patience is always sweet.'\n</thought>\n\nFinal Translation: The bowl of patience is always sweet."
    }
  ]
}

Intended Use

• Fine-tuning: Ideal for Supervised Fine-Tuning (SFT) of LLMs for Pashto reasoning.
• Translation Quality: Improves the handling of Pashto metaphors and proverbs.
• Educational: A resource for computational linguistics in the Pashto language.

Project Context

Developed by Nassim Nasibullah (nassimjp) in Japan as part of the iPashto.ai project. This dataset aims to bridge the gap in high-quality, reasoned datasets for low-resource languages like Pashto.

License

This dataset is released under the Apache License 2.0.