| --- |
| license: cc-by-4.0 |
| language: |
| - ar |
| tags: |
| - arabic |
| - dialect |
| - jordanian |
| - levantine |
| - safety |
| - function-calling |
| pretty_name: Jordanian Dialect Sample v1 |
| task_categories: |
| - text-generation |
| --- |
| # Jordanian Dialect Sample (v1) |
|
|
| **Levant AI** is building dialect-authentic Arabic training data for the Levantine region, starting with Jordanian/Shami dialect — collected natively, not translated from Modern Standard Arabic or English. |
|
|
| This is our first public sample: **30 examples** spanning three categories that reflect real gaps in current Arabic AI training data: |
|
|
| ## Categories |
|
|
| - **`general_conversation`** (10 examples) — everyday natural Jordanian dialect exchanges |
| - **`function_calling`** (10 examples) — dialect phrasing with clear actionable intent (booking, cancellation, price inquiry, order tracking), addressing the documented accuracy gap in Arabic function-calling |
| - **`safety_redteaming`** (10 examples) — realistic dialect-based prompt injection and social engineering attempts, addressing the documented gap where translated safety guardrails fail on native dialect input |
| |
| ## Format |
| |
| JSONL, one object per line with fields: category, dialect, text_arabic, text_english_gloss, notes |
| |
| ## Why this matters |
| |
| Most Arabic NLP datasets are either Modern Standard Arabic or machine-translated from English, which fails to capture how people actually speak — and fails specifically in safety and tool-use scenarios where dialect nuance changes model behavior. |
| |
| This dataset is manually written and reviewed by a native Jordanian speaker, not machine-translated. |
| |
| ## Roadmap |
| |
| This is v1 (sample). Planned expansion: 200-500+ examples per category, additional Levantine sub-dialects, and a living/maintained update cycle. |
| |
| ## Contact |
| |
| Built by LevantData. Open to collaboration, feedback, and partnership inquiries. |
