FatimahEmadEldin/Synanthic-Arabic-jokes-llama-3.3-70b-versatile

Dataset Card for Synanthic-Arabic-jokes-60k

Dataset Description

This dataset contains a large collection of synthetically generated Arabic jokes. The dataset was designed to capture a wide variety of Arabic dialects, topics, and comedic styles while maintaining strict content safety guidelines.

Data Generation

The jokes were generated using a programmatic loop that randomly combined a specific topic, dialect, and style into a prompt. The generation process explicitly prohibited politics, religion, racism, or profanity. All outputs were constrained to be short (one or two sentences).

Models Used

The dataset was synthetically generated using two Large Language Models:

• —Llama 3.3 70B Versatile (llama-3.3-70b-versatile) via the Groq API.

Diversity Matrix

The dataset prompts were built using a diverse matrix of attributes to ensure variety:

Topics (المواضيع)

• —Engineering, Medicine, Teaching, Students, Married Couples, Misers, Fools, Stoners, Football Players, Managers, Employees, Driving, Restaurants, Weather/Heat, Traffic, High Prices, Travel, Internet, Mothers, Fathers, Neighbors, Dieting, Weddings, Stock Market, Sleep, Laziness, Work/School Hours, Salary, Mobile Phones, and Friends.

Dialects (اللهجات)

• —General Saudi (باللهجة السعودية العامة).
• —Qassimi (باللهجة القصيمية).
• —Hijazi (باللهجة الحجازية).
• —Najdi (باللهجة النجدية).
• —Egyptian (باللهجة المصرية).
• —Levantine (باللهجة الشامية).
• —Simplified Modern Standard Arabic (باللغة العربية الفصحى المبسطة).
• —Gulf (باللهجة الخليجية).
• —Saudi Youth Style (بأسلوب شبابي سعودي).
• —Southern Saudi (باللهجة الجنوبية).

Styles (الأساليب)

• —Short and fast joke (نكتة قصيرة وسريعة).
• —Funny dialogue between two people (حوار مضحك بين شخصين).
• —Sarcastic Q&A (سؤال وجواب ساخر).
• —Embarrassing and funny situation (موقف محرج ومضحك).
• —Exaggerated description (وصف مبالغ فيه).

Limitations and Biases

Because the dataset is synthetically generated by LLMs, some jokes may rely on stereotypical tropes present in the training data of the models used. Although prompts strictly forbade offensive content, synthetic text may occasionally produce unpredictable outputs.