nassimjp/pashto_open-perfectblend

📘 Pashto Open‑PerfectBlend — Professional Dataset Card

(English + Pashto)

---

🏷️ Dataset Summary

Pashto Open‑PerfectBlend is a high‑fidelity Pashto translation of the Open‑PerfectBlend instruction dataset introduced in "Perfect Blend: Redefining RLHF with Mixture‑of‑Judges."

It contains chat, math, code, and instruction‑following data suitable for Pashto LLM pretraining, supervised fine‑tuning (SFT), RLHF pipelines, and general Pashto AI research.

🎨 خلاص‑پرفیکټ بلینډ د لارښوونې ډیټاسیټ یوه لوړ کیفیت لرونکې پښتو ژباړه ده چې په کامل مخلوط: د قاضیانو مخلوط سره د RLHF بیا تعریف کول مقاله کې معرفي شوې.

---

📦 Dataset Structure

Each record follows the OpenAI chat format in JSONL:

{
  "id": "000123",
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

هر ریکارډ په JSONL بڼه کې دی:

{
  "id": "000123",
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

---

🧠 Translation Pipeline

English

• Code, LaTeX, and structural syntax fully preserved
• Only natural language segments translated
• Hallucinated and boilerplate text removed
• Multi-threaded and resume-safe processing

پښتو

• کوډ، LaTeX او تخنیکي جوړښتونه بشپړ ساتل شوي
• یوازې طبیعي ژبې برخې ژباړل شوي
• تکراري او غلط متن پاک شوی
• څو تارې او بیا پیلیدونکې پروسې

---

🎯 Use Cases

English

• Pashto LLM pretraining
• Pashto SFT training
• Pashto RLHF / DPO experiments
• Pashto tokenizer development
• Educational and technical Pashto AI systems

پښتو

• د پښتو LLM روزنه
• د پښتو SFT روزنه
• د پښتو RLHF تجربې
• د پښتو tokenizer پراختیا
• تعلیمي او تخنیکي پښتو AI سیسټمونه

---

📏 Dataset Size

• ~46,000 records
• Size category: 10k–100k
• One of the largest publicly available Pashto instruction datasets

---

📁 Files

• data/pashto_messages_chunk_001.jsonl

---

📜 License

Apache-2.0

دا ډیټاسیټ د Apache-2.0 لایسنس لاندې خپور شوی دی.

---

🔗 Source & Credits

Original Dataset: Open‑PerfectBlend
Pashto Version: Nassim — Spinzar Enterprises Inc., Japan

---

📚 Citation

@dataset{pashto_open_perfectblend,
  author    = {Nassim},
  title     = {Pashto Open-PerfectBlend Dataset},
  year      = {2026},
  url       = {https://huggingface.co/datasets/nassimjp/pashto_open-perfectblend}
}

---

Made for the Pashto AI community ❤️