Mo-Abdalkader/Egyptian-Arabic-English-Parallel-Corpus

Egyptian Arabic-English Parallel Corpus

Author: Mohamed Abdalkader · LinkedIn · GitHub

A comprehensive Egyptian Arabic → English parallel corpus covering 1,800 topics from daily Egyptian life. Designed for fine-tuning large language models on Egyptian Arabic dialect translation and generation.

Dataset Structure

egyptian-arabic-english-parallel-corpus/
├── SFT/
│   ├── Train/
│   │   ├── topics/              # 1,800 individual topic JSON files
│   │   └── merged/              # Merged JSONL + CSV
│   └── Test/
│       ├── topics/              # 1,800 topic files (10 samples each)
│       └── merged/              # Merged JSONL + CSV
├── DPO/
│   ├── topics/                  # 1,800 topic files with preference pairs
│   └── merged/                  # Merged JSONL + CSV
├── metadata.json
├── dataset_infos.json
├── README.md
├── LICENSE
└── docs/
    └── dataset_statistics.md

Statistics

Split	Samples	Sentences	Description
SFT Train	277,319	554,638	Input → Output pairs
SFT Test	18,000	36,000	Input → Output pairs (10/topic)
DPO	124,681	498,724	Preference pairs (chosen/rejected)
Total	420,000	1,089,362

ID Format

Each sample has a unique ID: {split_id:02d}-{topic:05d}-{sample:04d}

Prefix	Split	Example
01	SFT Train	01-00001-0001 = Topic 1, Sample 1
02	SFT Test	02-01624-0010 = Topic 1624, Sample 10
03	DPO	03-01800-0050 = Topic 1800, Sample 50

Loading

Python

import json

# Load a specific topic
with open("SFT/Train/topics/0001-First-day-at-school.json") as f:
    data = json.load(f)

# Load merged file
with open("SFT/Train/merged/train.jsonl") as f:
    lines = [json.loads(line) for line in f]

Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("json", data_files={
    "train": "SFT/Train/merged/train.jsonl",
    "test": "SFT/Test/merged/test.jsonl"
})

Topic Files Format

Each topic file is a JSON array of samples:

[
  {
    "id": "01-00001-0001",
    "category": "First-day-at-school",
    "input": "...Egyptian Arabic...",
    "output": "...English translation..."
  }
]

DPO topic files have chosen_input, chosen_output, rejected_input, rejected_output instead.

License

CC BY 4.0 — see LICENSE for details.

Citation

@misc{abdalkader2026egyptian,
  author    = {Mohamed Abdalkader},
  title     = {Egyptian Arabic-English Parallel Corpus: A +1M-Sentence Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Mo-Abdalkader/Egyptian-Arabic-English-Parallel-Corpus}
}