Syrinesmati/tunisian-question-response-dataset

Tunisian Question-Response Dataset

Dataset Description

A supervised fine-tuning (SFT) dataset of ~31,000 question-answer pairs in Tunisian Arabic dialect (Darija). The dataset was designed to support the development of language models that can understand and generate natural Tunisian dialect, a historically under-resourced Arabic variety with strong French code-switching influence.

Pairs are written primarily in Arabic script, with natural mixing of Latin-script French terms where appropriate to the topic — reflecting how Tunisians actually write in informal and semi-formal digital contexts.

• Curated by: Syrine Smati
• Language: Tunisian Arabic (Darija) — ar-TN
• License: CC-BY 4.0

Dataset Structure

Each example contains three fields:

Field	Type	Description
instruction	string	A question or request in Tunisian dialect
response	string	A natural, helpful answer in Tunisian dialect
category	string	Thematic category (see below)

Categories

The dataset covers 9 thematic categories:

Category	Description
cultural_knowledge	Tunisian culture, traditions, food, music, proverbs, history
advice_guidance	Life advice, health, relationships, practical decisions
general_knowledge	Science, geography, sports, current events
task_oriented	Write, explain, summarize, translate, list
emotional_support	Emotional and psychological situations, encouragement
everyday_conversation	Greetings, daily life, casual chat
storytelling	Stories, folk tales, narrative requests
student_life	University, exams, scholarships, student problems
opinions_recommendations	Opinions, comparisons, recommendations

Data Splits

Split	Examples
Train	25,342
Test	6,336
Total	31,678

Dataset Creation

Curation Rationale

Tunisian Arabic is spoken by ~12 million people but is severely under-represented in NLP resources. Existing Arabic datasets use Modern Standard Arabic (MSA) or other dialects, which do not reflect how Tunisians communicate in practice. This dataset was created to enable fine-tuning of multilingual LLMs on authentic Tunisian dialect.

Source Data

Data Collection and Processing

Instruction-response pairs were AI-generated using a large language model, with each pair reviewed and validated by native Tunisian Arabic speakers. The curation process included:

• Dialect verification: Responses were scored on Tunisian dialect marker presence and penalized for MSA markers. Entries with no dialect markers or high MSA scores were flagged and reviewed or rewritten.
• Deduplication: Exact and near-duplicate instructions were removed.
• Quality filtering: Very short responses (< 5 words), truncated entries, and repetitive generations were removed.

Who are the source data producers?

Questions and responses were generated using a large language model and supervised by native speakers of Tunisian Arabic to ensure dialect authenticity, cultural accuracy, and response quality.

Annotations

No separate annotation layer. Dialect quality and category labels were assigned during generation and validated during the curation process by the dataset author.

Personal and Sensitive Information

The dataset does not contain personal information, private conversations, or sensitive data. All content is synthetic (AI-generated) and does not reference real individuals.

Bias, Risks, and Limitations

• Dialect variation: Tunisian Arabic has regional variation (Tunis, Sfax, Sousse, Sahel, etc.). The dataset reflects primarily urban Tunisian dialect and may not fully represent rural or regional varieties.
• AI generation artifacts: Despite human review, some responses may retain subtle MSA influences or unnatural phrasing not caught during curation.
• Category imbalance: cultural_knowledge and advice_guidance are more represented than storytelling, student_life, and opinions_recommendations.
• Code-switching scope: French code-switching is present but not exhaustively represented across all domains.

Recommendations

This dataset is intended for use in supervised fine-tuning of language models for Tunisian dialect understanding and generation. Users should evaluate model outputs with native speakers before deployment. It should not be used as a factual knowledge base.

Citation

If you use this dataset, please cite:

@dataset{smati2025tunisian,
  author    = {Smati, Syrine},
  title     = {Tunisian Question-Response Dataset},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Syrinesmati/tunisian-question-response-dataset}
}

Dataset Card Contact

For questions or feedback, please open a discussion on the dataset page.