opus-100-ps-en

A high‑quality Pashto → English parallel corpus reverse‑engineered from the original OPUS‑100 (en‑ps) dataset by Helsinki‑NLP.
This dataset enables robust Pashto→English machine translation, Pashto language understanding, auto‑detection, and dual‑direction translation engines for modern NLP systems.

---

🌍 Motivation

Pashto is a low‑resource language in global NLP research. While OPUS‑100 provides English→Pashto pairs, the reverse direction (Pashto→English) is essential for:

• Pashto language modeling
• Pashto semantic reasoning
• Pashto auto‑detection
• Bidirectional MT systems
• Pashto linguistic analysis
• Pashto AI research at scale

This dataset fills that gap by providing a clean, reversed ps‑en corpus.

---

📦 Dataset Structure

The dataset contains three Parquet files:

train-00000-of-00001.parquet
validation-00000-of-00001.parquet
test-00000-of-00001.parquet

Each file includes two columns:

• ps — Pashto sentence
• en — English translation

Example

ps: "زه ښوونځي ته ځم"
en: "I am going to school"

---

🏗️ Technical Architecture

1. Source Extraction Layer

The original OPUS‑100 en‑ps dataset is loaded from HuggingFace in Parquet format.
Each record contains:

{
  "translation": {
    "en": "...",
    "ps": "..."
  }
}

2. Reverse Engineering Layer

The dataset is reconstructed by swapping the direction:

• Source → ps
• Target → en

Extraction logic:

df_reversed = pd.DataFrame({
    "ps": df["translation"].apply(lambda x: x["ps"]),
    "en": df["translation"].apply(lambda x: x["en"])
})

3. Split Preservation

The original OPUS‑100 splits are preserved:

• Train
• Validation
• Test

Ensuring reproducibility and MT evaluation consistency.

4. Storage Layer

All splits are saved in Parquet format for:

• Fast loading
• Efficient storage
• Compatibility with HF Datasets, PyTorch, TensorFlow, Spark

5. Schema Definition

Column	Type	Description
ps	string	Pashto source sentence
en	string	English target translation

6. Compatibility Layer

Fully compatible with:

• HuggingFace Datasets
• Transformers MT pipelines
• PyTorch DataLoader
• TensorFlow tf.data
• Spark distributed processing
• SentencePiece / BPE tokenizer training

7. Downstream System Design

Intended for:

• Pashto→English MT engines
• Dual en↔ps translation systems
• Pashto auto‑detect modules
• Pashto semantic reasoning
• Pashto LLM fine‑tuning
• Pashto linguistic evaluation

---

🧠 Use Cases

Machine Translation

• Pashto→English MT model training
• Dual‑direction MT engines (en↔ps)
• Fine‑tuning multilingual LLMs

Pashto NLP Research

• Pashto semantic embeddings
• Pashto syntax & morphology analysis
• Pashto text normalization

AI Systems

• Pashto auto‑detect
• Pashto reasoning modules
• Pashto linguistic evaluation

---

📊 Dataset Statistics

Split	Approx. Size	Format
Train	~1.8M pairs	Parquet
Validation	~2K pairs	Parquet
Test	~2K pairs	Parquet

(Counts may vary slightly depending on OPUS filtering.)

---

🏷️ Tags

• pashto
• pashto-nlp
• pashto-dataset
• translation
• machine-translation
• opus100
• bilingual-corpus
• parallel-corpus
• ps-en
• low-resource-language
• nlp
• linguistics
• ai-datasets

---

📚 Source

Derived from:

Helsinki‑NLP / OPUS‑100
Language pair: en-ps
Reversed to: ps-en

---

📜 License

This dataset follows the licensing terms of the original OPUS‑100 dataset.

---

🙏 Acknowledgements

Special thanks to Helsinki‑NLP for creating OPUS‑100 and enabling multilingual NLP research.

---

✏️ Citation

If you use this dataset, please cite OPUS‑100:

Tiedemann, J., & Thottingal, S. (2020). OPUS‑100: A Multilingual Corpus for Translation.

---

## 🔗 Useful Links

- **[OPUS‑100 Original Dataset](https://huggingface.co/datasets/Helsinki-NLP/opus-100)**  
- **Pashto NLP Resources**  
- **Model Training Guide**  

---