DziriAlign: Algerian Darija & Cultural Preference Dataset

DziriAlign is a preference alignment dataset containing 1,000 high-quality samples designed to align Large Language Models (LLMs) with Algerian Arabic (Darija) language, sociocultural norms, bargaining etiquette, and humor.

It is formatted as a preference dataset (prompt, chosen, rejected) ideal for Direct Preference Optimization (DPO), Reinforcement Learning from Human Feedback (RLHF), or supervised fine-tuning (SFT) evaluations.

---

Dataset Overview

Standard multilingual and Arabic language models often struggle to capture the nuances of Algerian Darija due to its unique blend of Arabic, French, Berber (Tamazight), and localized slang. Furthermore, standard models often reply in overly formal Modern Standard Arabic (MSA) or present culturally mismatched advice.

DziriAlign solves this by providing pairs of chosen (culturally/linguistically authentic and natural Algerian) and rejected (robotic, overly formal, or culturally tone-deaf) responses.

---

Dataset Composition

The dataset is a compilation of four specialized subsets of 250 items each, covering key aspects of Algerian linguistic and social life:

Sub-dataset Name	Size	Focus / Context	Description
Algerian Market Negotiation	250 pairs	Local commerce & bargaining	Features typical market negotiations (buying clothes, electronics, repair services). chosen uses local bargaining logic and friendly customer terms; rejected uses dry, corporate, and fixed-price responses.
Algerian Socio-Linguistic	250 pairs	Daily greetings & social interactions	Covers condolences, greetings, congratulations, and everyday hospitality. chosen uses warm, authentic Darija expressions; rejected uses generic, formal MSA or robotic replies.
Algerian Sociocultural	250 pairs	Societal norms & family advice	Addresses family advice, independent living, career issues, and local expectations. chosen gives nuanced, culturally balanced advice; rejected gives aggressive, Westernized or rigid advice that ignores local realities.
Generating Algerian Humor	250 pairs	Sarcasm, jokes, and local wit	Captures local Algerian humor and memes (e.g. slow internet, heatwaves). chosen plays along with authentic humor; rejected offers sterile technical support or literal interpretations.

---

Schema & Format

The dataset is saved in strict JSON Lines (JSONL) format, with each line containing a single JSON object with the following fields:

• prompt (string): The input instruction, question, or scenario (primarily in Algerian Darija, Arabizi, or mixed Arabic script).
• chosen (string): The preferred response, representing natural Algerian dialect, local norms, humor, or negotiation style.
• rejected (string): The non-preferred response, containing dry/formal MSA, overly literal translations, or culturally disconnected advice.

---

Data Examples

Here are representative examples from the combined dataset:

1. Market Negotiation

{
  "prompt": "Khoya, had sabat chhal mellekher? Ghali chwiya.",
  "rejected": "The listed price is final and non-negotiable as per our store policy. Would you like to proceed to checkout?",
  "chosen": "Wlh ya khouya ghir rani nbi3 fih b ras malou, mais ma ykoun ghir khatek, nsa3dek fih chwiya, het 4000 w bsahtek."
}

2. Socio-Linguistic

{
  "prompt": "صحا فطورك.",
  "rejected": "شكرا لك، أتمنى لك إفطارا شهيا.",
  "chosen": "يسلمك، صحا فطورك نتا تاني، ربي يتقبل منا ومنكم صالح الأعمال."
}

3. Sociocultural

{
  "prompt": "يا خويا عييت، راني خدامة وندخل دراهمي بصح دارنا يعاملوني كي الطفلة. راني نخمم نكري وحدي ونستقل، بصح خايفة من الهدرة.",
  "rejected": "أنت مستقلة مالياً وتتمتعين بحقوقك الكاملة. ضعي حدوداً صارمة مع عائلتك وانتقلي للعيش بمفردك من أجل صحتك النفسية. لا تبالي بآراء المجتمع.",
  "chosen": "يا أختي، الاستقلالية مليحة، بصح في مجتمعنا طفلة تسكن وحدها بلا سبة قاوية تجيب الهدرة وتدير مشاكل مع العايلة بلا فايدة. حاولي تفاهمي معاهم بالعقل وصارحيهم بلي راكي تكبري وتتحملي مسؤولية، بلا ما تخسري رضاهم."
}

4. Algerian Humor

{
  "prompt": "Connexion fel dzair ki selhoufa, rak teb3at message yelhaq fel 2030.",
  "rejected": "أعتذر عن تجربتك السيئة مع الإنترنت. يمكنك محاولة إعادة تشغيل جهاز التوجيه الخاص بك أو الاتصال بمزود الخدمة للحصول على مساعدة فنية لتحسين السرعة.",
  "chosen": "Wlh ghir sdaqte! Dir kima hna, eb3at l'message w roh tayeb qahwa w zid dir sieste, balak yelhaq ki tnoud."
}

---

Technical Information & Generation

• Total Count: 1,000 entries (250 items per subset)
• Encoding: UTF-8
• File Format: Strict JSON Lines (one JSON object per line)

How to load the dataset in Python:

import json

dataset_path = "DziriAlign.jsonl"

with open(dataset_path, "r", encoding="utf-8") as f:
    for line in f:
        item = json.loads(line)
        # Process the item: item["prompt"], item["chosen"], item["rejected"]
        print(f"Prompt: {item['prompt']}")
        print(f"Chosen: {item['chosen']}")
        print(f"Rejected: {item['rejected']}")
        break

Loading with Hugging Face datasets:

You can load the dataset directly from Hugging Face Hub:

from datasets import load_dataset

dataset = load_dataset("touati-kamel/DziriAlign")
print(dataset["train"][0])

License

This dataset is made available under the MIT License. You are free to use, modify, and distribute it for academic research or application development, provided that the original source is cited.