nassimjp/pold-cpt-corpus

Pashto CPT Corpus — POLD (Shuffled)

This dataset contains raw, unfiltered Pashto text extracted from the original
zirak-ai/pold dataset, which is licensed under Creative Commons Attribution–NonCommercial 4.0 (CC BY‑NC 4.0).

Only the text field was extracted and shuffled to prepare it for
Continued Pretraining (CPT) of Pashto language models.

📦 File Included

pold_shuffled.jsonl

Each line is a JSON object:

{"text": "…Pashto text…"}

---

🙏 Acknowledgements & Community Credits

This dataset exists because of many years of work by Pashto digital volunteers,
open‑source contributors, and linguistically skilled individuals who preserved, cleaned,
and organized Pashto text long before modern LLMs existed.

We acknowledge:

• The original authors of the POLD dataset
• The Pashto AI community, whose early efforts built the foundation for Pashto NLP
• The highly educated Pashto linguists, teachers, journalists, and researchers
  who contributed to the digital Pashto ecosystem
• All individuals who worked—often unpaid—to bring Pashto language resources online

Without their dedication, Pashto AI would not be possible today.

---

🎯 Purpose of This Release

This dataset is intended for:

• Continued Pretraining (CPT)
• Tokenizer adaptation
• Pashto language modeling research
• Dialect and slang modeling
• Cultural‑aware Pashto NLP systems

It preserves:

• slang
• dialects
• spelling variations
• emotional expressions
• humor and sarcasm
• political and social commentary
• natural Pashto internet language

This authenticity is essential for realistic Pashto LLMs.

---

⚠️ Ethical & Safety Notes

This dataset contains:

• strong language
• insults
• political opinions
• religious expressions
• emotionally charged content
• references to harmful or extremist groups in user‑generated context

These references reflect real social‑media discourse, not endorsement.
Violent extremist organizations (e.g., ISIS/Daesh, Al‑Qaeda, the Taliban) are responsible
for severe harm and human rights violations. Any mentions in the dataset come from
public user‑generated text and do not represent support.

Users must ensure:

• responsible use
• compliance with local laws
• proper safety alignment during model training

---

📉 Limitations

• Contains noise, typos, and inconsistent spelling
• Not balanced across dialects
• Not suitable for supervised tasks without preprocessing
• May include culturally sensitive or offensive content

---

👤 Maintainer

nassimjp

Dataset prepared and released for the advancement of Pashto NLP research.