nassimjp/pashto_wheat_sft_14k

🌾 Pashto Wheat SFT 14K Dataset

📌 Introduction

Pashto Wheat SFT 14K is a high-quality, lightweight Supervised Fine-Tuning (SFT) dataset tailored specifically for training native Pashto AI models. The dataset comprises roughly 14,000 short, concise, and highly informative instructional examples natively adapted to cover essential tech concepts, programming languages, and general knowledge structures.

The philosophy behind this dataset reflects a local Pashto proverb: *"غنم پخېږي" (The wheat is ripening)*—signifying the dawn of digital sovereignty and independent foundational infrastructure for the Pashto language in the era of Large Language Models (LLMs).

🛠️ Key Features

• Tech-Keyword Preservation: Advanced filtering was applied during curation to shield critical software engineering terms (e.g., React, Python, Docker, API, Git) from being awkwardly translated into literal prose, ensuring the model retains exact structural and technical proficiency.
• Optimized for SFT: Short-context sequences reduce compute costs and memory footprint while teaching models to provide precise, direct, and zero-fluff responses.
• Standardized Schema: Packaged cleanly in the widely accepted {"prompt": "...", "response": "..."} format.

---

📊 Dataset Structure

Each record contains a instruction-response pair wrapped in a clean JSON Lines (JSONL) structure:

{
  "prompt": "React څه شی دی؟",
  "response": "React یو JavaScript کتابتون دی چې د کارن انٹرفیس جوړولو لپاره کارول کیږي."
}

Data Fields

• prompt: The user instruction or technical inquiry written in clean, naturally flowing Pashto text.
• response: The precise, direct ground-truth answer in Pashto, maintaining necessary English coding/tech nomenclature where required.

---

🚀 Intended Use Cases

• Instruction Tuning (SFT): Perfect for fine-tuning baseline models (such as Llama, Mistral, or Qwen) to specialize in or understand the Pashto language structure.
• Vocabulary Expansion: Enhances localized computing vernacular for low-resource NLP initiatives.
• Tech Chatbots: Training conversational agents capable of answering coding and IT-related queries directly in Pashto.

📄 License

This dataset is openly shared under the Apache 2.0 License, promoting collaboration and community-driven AI advancements for the Pashto language.

---

Created with 🌾 and passion by the Pashto-AI Initiative (nassimjp). Let's build the future of native Pashto computing together!