Datasets:

HeshamHaroon
/

arabic-dialect-dpo

language:
  - ar
license: apache-2.0
size_categories:
  - 10K<n<100K
task_categories:
  - text-generation
tags:
  - dpo
  - rlhf
  - arabic
  - egyptian-arabic
  - saudi-arabic
  - dialect
  - preference
  - alignment
  - orpo
  - grpo
pretty_name: Arabic Dialect DPO - Egyptian & Saudi
dataset_info:
  - config_name: egyptian
    features:
      - name: prompt
        dtype: string
      - name: chosen
        dtype: string
      - name: rejected
        dtype: string
      - name: category
        dtype: string
      - name: language
        dtype: string
      - name: dialect
        dtype: string
    splits:
      - name: train
        num_examples: 11038
  - config_name: saudi
    features:
      - name: prompt
        dtype: string
      - name: chosen
        dtype: string
      - name: rejected
        dtype: string
      - name: category
        dtype: string
      - name: language
        dtype: string
      - name: dialect
        dtype: string
    splits:
      - name: train
        num_examples: 11500
configs:
  - config_name: egyptian
    data_files:
      - split: train
        path: egyptian/train-*
  - config_name: saudi
    data_files:
      - split: train
        path: saudi/train-*

Arabic Dialect DPO Dataset - Egyptian & Saudi

The first large-scale Arabic dialect preference dataset for DPO/ORPO/GRPO alignment training. Contains 22,538 preference triples across two major Arabic dialects: Egyptian (Masry) and Saudi (Najdi).

Dataset Summary

Config	Dialect	Rows	Language Code
`egyptian`	Egyptian Arabic (مصري)	11,038	ar-EG
`saudi`	Saudi Arabic (سعودي نجدي)	11,500	ar-SA
Total		22,538

Usage

from datasets import load_dataset

# Load Egyptian Arabic
egyptian = load_dataset("HeshamHaroon/arabic-dialect-dpo", "egyptian")

# Load Saudi Arabic
saudi = load_dataset("HeshamHaroon/arabic-dialect-dpo", "saudi")

# Load both
egyptian_train = load_dataset("HeshamHaroon/arabic-dialect-dpo", "egyptian", split="train")
saudi_train = load_dataset("HeshamHaroon/arabic-dialect-dpo", "saudi", split="train")

Format

Each row contains a DPO preference triple:

Field	Description
`prompt`	User question/request in dialect Arabic
`chosen`	High-quality response in natural dialect (preferred)
`rejected`	Low-quality response in formal/generic Arabic (dispreferred)
`category`	Topic category
`language`	Language code (ar-EG or ar-SA)
`dialect`	Dialect name

Examples

Egyptian Arabic

Prompt: ازاي أعمل CV كويس يلفت نظر الشركات؟
Chosen: أول حاجة لازم تعملها إن الـ CV يكون منظم وواضح. خليه صفحة واحدة أو اتنين بالكتير...
Rejected: يمكنك إنشاء سيرة ذاتية احترافية من خلال اتباع بعض الإرشادات العامة...

Saudi Arabic

Prompt: وش أحسن طريقة أطور فيها مهاراتي عشان رؤية 2030؟
Chosen: يالله أبشرك، رؤية 2030 فتحت أبواب كثيرة للتطوير. أول شيء حدد المجال اللي يناسبك...
Rejected: يمكنك تطوير مهاراتك من خلال البحث عن الدورات التدريبية المتاحة عبر الإنترنت...

Key Contrast

The preference signal is based on:

Chosen: Natural dialect, culturally relevant, detailed, practical, locally grounded
Rejected: Formal Arabic (فصحى), generic, short, culturally disconnected, "search the internet" style

Generation Details

Aspect	Detail
Generation	Synthetic via large language models
Deduplication	Exact prompt matching
Quality control	Length filtering + dialect authenticity checks

Use Cases

DPO training for Arabic dialect models
ORPO/GRPO alignment for Egyptian or Saudi Arabic
Reward model training with clear preference signals
Arabic dialect research comparing Egyptian vs Saudi
Cultural NLP studying how dialect affects AI helpfulness

Limitations

Synthetically generated (not human conversations)
May not capture all subdialects within Egyptian/Saudi Arabic
Rejected responses are intentionally generic, not adversarial

Citation

@dataset{arabic_dialect_dpo_2026,
  title={Arabic Dialect DPO Dataset - Egyptian & Saudi},
  author={Hesham Haroon},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/HeshamHaroon/arabic-dialect-dpo}
}

License

Apache 2.0