nassimjp/pashto-cpt

Pashto CPT — Clean Pretraining Corpus

This dataset contains a cleaned, normalized, and noise‑reduced Pashto text corpus designed specifically for CPT (Continued Pretraining) of Pashto Large Language Models.

It is built from diverse Pashto sources and processed with a custom cleaning pipeline to remove:

• Urdu punctuation noise (۔۔۔, ۔۔۔۔)
• Mixed Latin/Urdu dot clusters
• Repeated commas (،،،)
• Non‑Pashto symbols
• Spacing artifacts
• Encoding inconsistencies

The result is a high‑quality, tokenizer‑friendly Pashto corpus suitable for LLM pretraining and domain adaptation.

---

📦 Contents

• pashto_corpus_output_clean.jsonl

• Each line is a JSON object:

  {"text": "ماشومان په ګډه سندرې وایي"}
  

• Fully UTF‑8 normalized

• No metadata, no labels — pure text for pretraining

---

🎯 Intended Use

• Continued pretraining (CPT)
• Pashto‑specific LLM development
• Tokenizer training
• Embedding model training
• Language modeling research

---

🧹 Cleaning Pipeline Summary

The corpus was processed with a custom Pashto cleaner that:

• Normalizes Urdu + Latin dot clusters → ۔
• Normalizes commas → ،
• Removes Arabic noise characters (ـ, …, ؛)
• Removes non‑Pashto symbols
• Collapses multiple spaces
• Preserves Pashto diacritics and letters

This ensures stable training, better tokenization, and higher model quality.

---

📜 License

Released for research and model development.
Users are responsible for compliance with local laws and platform policies.

---

👤 Author

nassimjp

---