💎 Pashto-SFT-263k-Cleaned

A highly clean, structurally verified, and randomized Supervised Fine-Tuning (SFT) dataset containing 263,690 high-quality instructional pairs in the Pashto language (ps).

This dataset was compiled, aligned, and optimized as part of Project Talanda-TM (Train J) to advance natural language processing, instruction-following capabilities, and generative AI alignment for the Pashto language.

---

📊 Dataset Summary

• Total Records: 263,690 conversational samples
• Format: ShareGPT / OpenAI Messages Format (jsonl)
• Language: Pashto (ps)
• Components: Balanced blend of creative writing, multi-turn reasoning, scientific concepts, and instruction-following.
• Randomization: Fully shuffled with a fixed random seed (42) to distribute topics evenly and prevent training cross-entropy bias.

---

📁 Dataset Structure

Each row in the dataset consists of a single JSON object containing a "messages" list, perfectly optimized for modern SFT engines (like Hugging Face SFTTrainer, Axolotl, or LLaMA-Factory).

Data Schema

{
  "messages": [
    {
      "role": "user",
      "content": "User prompt or instruction in Pashto"
    },
    {
      "role": "assistant",
      "content": "Aligned, high-quality response in Pashto"
    }
  ]
}

Raw Format Preview

Example 1: Creative Writing / Multi-Turn Fiction

{
  "messages": [
    {"role": "user", "content": "ماته د یو جادوګر په اړه یوه لنډه کیسه ولیکئ چې د ځواکمن وجود د رابللو لپاره رسم ترسره کوي..."},
    {"role": "assistant", "content": "تیاره څراغ کوټه د بخور له بوی سره ډکه وه او د شمعې څپې په دیوالونو کې عجیب سیوري راوتلې. جادوګر مالکم ویچ ووډ..."}
  ]
}

Example 2: Scientific & Academic Reasoning

{
  "messages": [
    {"role": "user", "content": "د ستورو په شاوخوا کې د سوپرنووا اغیزې څه دي؟"},
    {"role": "assistant", "content": "یو سوپرنووا، د ستوري ناورین چاودنه، په شاوخوا کې د ستوري منځني (ISM) په اړه ژوره اغیزه لري...\n1. **د شاک څپې او د چاودنې څپې**..."}
  ]
}

---

🛠️ Data Engineering & Quality Controls

To ensure high structural integrity, the dataset went through rigorous assembly phases:

1. Alphanumeric Alignment: Extracted and mapped matching base IDs across disjointed raw translation outputs, solving alignment bugs on complex multi-chunk configurations.
2. Graceful Error Resilience: Safely dropped unaligned text chunks while retaining over 99.6% of the original sample footprint.
3. Reproducible Shuffle: Mixed completely using Python and Unix native stream components under Seed 42.

---

🚀 Quick Start / Usage

Using Hugging Face Datasets Library

from datasets import load_dataset

# Load this dataset directly from the hub
dataset = load_dataset("nassimjp/Pashto-SFT-263k-Cleaned")

# Inspect the first training instance
print(dataset['train'][0])

Intended Use Cases

• Supervised Fine-Tuning (SFT): Aligning raw base LLMs (such as LLaMA-3, Qwen-2.5, or Mistral) to fluently converse and follow complex instructions in Pashto.
• Cross-Lingual Adaptation: Vocabulary and contextual extension tasks targeting Middle Eastern and South Asian regional language support.

---

⚖️ License & Attribution

This dataset is distributed under the Apache-2.0 License. You are free to modify, distribute, and utilize this dataset for commercial and non-commercial model training runs.