Persian-Summarization-Large 📚

The largest unified Persian text summarization dataset for LLM and NLP research.

Persian-Summarization-Large merges, normalizes, and deduplicates multiple prominent open-source Persian summarization corpora into a single, research-ready .parquet dataset — eliminating preprocessing overhead for researchers and practitioners.

---

Dataset Summary

Property	Value
Language	Persian (fa)
Task	Abstractive Text Summarization
Format	Apache Parquet
License	CC-BY-NC-4.0
Size	100K–1M samples

---

Why Use This Dataset?

• Largest merged Persian summarization corpus on Hugging Face
• Covers news, encyclopedic, religious, and web domains for robust generalization
• Suitable for fine-tuning mT5, mBART, ParsBERT, and multilingual seq2seq models
• Fully cleaned and normalized — no preprocessing required
• One-line loading via 🤗 datasets

---

Dataset Structure

Each row represents one document–summary pair:

Column	Type	Description
source	string	Origin corpus (e.g., xlsum_fa, pn_summary, tebyan, wiki_summary)
text	string	Full input Persian document
summary	string	Target summary

---

🛠️ How to Use

from datasets import load_dataset

dataset = load_dataset("Arshia82sbn/Persian-Summarization-Large")
print(dataset["train"][0])

Filter by source corpus:

python
news = dataset["train"].filter(lambda x: x["source"] == "pn_summary")

---

## Data Processing Pipeline

1. **Format Normalization** — Unified varying column names across all sources into standard `text` / `summary` schema.
2. **Text Cleaning** — Converted structural tags (e.g., `<br>`) to newlines; normalized whitespace and Unicode.
3. **Quality Filtering** — Removed `NaN` values, documents shorter than 20 characters, and summaries shorter than 5 characters.
4. **Efficient Export** — Saved as `.parquet` for fast streaming and loading.

---

## Source Corpora

| Corpus | Domain |
|---|---|
| XL-Sum (Persian) | News (BBC Persian) |
| PN-Summary | Persian news |
| Tebyan | Articles & religious/cultural content |
| WikiSummary (fa) | Persian Wikipedia |

---

## ⚖️ Attribution & Copyright

This dataset is a derivative work. All credit belongs to the original authors. **You must cite the original sources if you use this dataset.**

**XL-Sum (Persian)**
> Hasan, T. et al. "XL-Sum: Large-Scale Multilingual Abstractive Summarization for 44 Languages." *Findings of ACL-IJCNLP 2021*.

**PN-Summary**
> Please attribute the original creators of the PN-Summary Persian news summarization dataset.

**Tebyan**
> Please attribute the original creators of the Tebyan dataset collection.

**WikiSummary (Persian)**
> Summaries derived from Persian Wikipedia dumps.

---

## License

Distributed under [`CC-BY-NC-4.0`](https://creativecommons.org/licenses/by-nc/4.0/) to respect the non-commercial and attribution constraints of the underlying source datasets. **Commercial use is not permitted.**

---

## Citation

If you use this dataset, please cite the original corpora listed above and link back to this repository.