nassimjp/english_pashto_parallel-corpus

English–Pashto Parallel Corpus

A high-quality English–Pashto parallel corpus designed for machine translation, supervised fine-tuning, Pashto LLM training, and corpus quality research.

---

📁 Dataset Structure

1. translation_clean.jsonl

High-quality parallel data suitable for:

• Machine Translation
• Supervised Fine-Tuning (SFT)
• Pashto LLM Training
• Instruction Tuning
• Reasoning Alignment

Each record follows:

{
  "id": 12345,
  "source": "English sentence...",
  "target": "Pashto translation..."
}

---

2. translation_anomalies.jsonl

Contains records identified during automatic anomaly detection.

Typical anomaly categories include:

• lexical glitches
• repeated tokens
• mistranslations
• English leakage
• malformed formatting
• hallucinated Pashto

This split is valuable for:

• corpus debugging
• automatic correction pipelines
• quality benchmarking
• anomaly detection research

---

📊 Dataset Statistics

Metric	Value
Total records	830,879
Clean records	713,429
Anomalies	117,450
Clean ratio	85.86%
Anomaly ratio	14.14%

---

🧠 Use Cases

• English ↔ Pashto Machine Translation
• Pashto LLM Pretraining
• Supervised Fine-Tuning (SFT)
• Instruction Tuning
• Reasoning Alignment
• Corpus Quality Research
• Anomaly Detection Benchmarking
• Data Cleaning Pipeline Development

---

🧪 Data Cleaning Pipeline

The corpus was processed through a multi-stage quality pipeline including:

• JSON merge validation
• Translation cache merging
• Deduplication
• Pashto extraction
• Length-based filtering
• Token repetition detection
• English leakage detection
• Lexical anomaly detection
• Quality scoring
• Clean vs anomaly dataset splitting

This workflow is designed to produce reproducible, high-quality datasets for Pashto NLP and LLM development.

---

👤 Author

Nassim Nasibullah

Spinzar Enterprises Inc. (Yūgen Kaisha), Japan

---

📚 Citation

@dataset{english_pashto_parallel_corpus_2026,
  author       = {Nassim Nasibullah},
  title        = {English--Pashto Parallel Corpus},
  year         = {2026},
  publisher    = {Hugging Face},
  journal      = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/nassimjp/english_pashto_parallel-corpus}
}

---

🚀 Project Goal

This dataset is part of the broader effort to build open, reproducible infrastructure for Pashto NLP, including machine translation, large language model training, instruction datasets, and quality evaluation pipelines.