اختبر أي موديل ذكاء اصطناعي على اللهجة الإماراتية: 12 برومبت وجدول تقييمTest Any AI Model on Emirati Arabic: 12 Prompts and a Scoring Sheet

Key terms

Falcon-Emirati

A 7B language model from the Technology Innovation Institute (TII) in Abu Dhabi, built on Falcon-H1-Arabic and trained on native Emirati content. You can try it on Falcon Chat.

Alyah

An Emirati Arabic benchmark covering everyday language, figurative expressions, heritage knowledge and poetry. TII reports Falcon-Emirati at 84.83% on it.

Benchmark

A fixed set of questions you give several models so you can compare them fairly. Here you build a small one of your own.

Model

The AI that answers you. Each model learned from different data, so its grasp of dialect differs.

The steps

1. 1

   Step 1

   Pick the models to compare

   Pick two or three. For example Falcon-Emirati on Falcon Chat, and the model you use every day.

   Open each one in a fresh chat with no prior instructions, so the comparison is fair.

   Tip: Falcon-Emirati is at chat.falconllm.tii.ae, choose it from the model list.

2. 2

   Step 2

   Prepare 12 questions in 4 groups

   Split the questions the way Alyah does: everyday language, proverbs and figures of speech, heritage, and poetry. Three questions each.

   Write them in your own dialect, not in formal Arabic. That is the real test.

   PromptSample questions per group (in Arabic)

   Copy

   كلام يومي: شو معنى «يرمس»؟ واستخدمها في جملة.
   كلام يومي: رد على عميل يقول «الطلبية تأخرت وايد» بأسلوب ودود وباللهجة.
   أمثال: اشرح لي مثل «اللي ما يعرف الصقر يشويه» ومتى نقوله.
   تراث: شو هو «البرجيل» في البيوت الإماراتية القديمة؟ اشرحه لطفل.
   شعر: اكتب بيتين نبطيين عن [your topic].

3. 3

   Step 3

   Give every model the exact same question

   Paste the same text into each model, word for word. Do not reword it when an answer disappoints you.

   Save the answers in a table: one column per model, one row per question.

   PromptFixed line before every question

   Copy

   رد علي باللهجة الإماراتية، بدون فصحى، وبجملتين أو ثلاث بس.
   
   [your question]

4. 4

   Step 4

   Score each answer from 0 to 2

   Zero: it did not understand or replied in formal Arabic. One: it understood but the dialect is weak or wrong. Two: it understood and answered like a local.

   The total is out of 24. Add a short note on each model's biggest mistake.

   Tip: Get a second native speaker to score with you. Alone, your eye gets used to the answers.

5. 5

   Step 5

   Let a model summarise the result

   When you finish, paste the table into Claude or any model and ask it to rank the results and give you the bottom line.

   PromptSummary prompt

   Copy

   This is a scoring table for AI models on Emirati Arabic. Each question scores 0 to 2.
   
   [paste the table]
   
   Rank the models by total and by group (everyday, proverbs, heritage, poetry). Then tell me in three points which model to use for [your use: customer replies, captions, ...] and why.

6. 6

   Step 6

   Decide by your use, not by size

   The biggest model is not always best at dialect. If your work is replying to local customers, the everyday group matters most. For cultural content, proverbs and heritage.

   Run the same test whenever a new model ships. Same questions, so you compare on the same scale.

Ready prompts to copy

واحد قال لي: «[a sentence in dialect]». شو قصده بالضبط؟ وشو الرد المناسب له باللهجة؟

حوّل هذا النص للهجة الإماراتية مثل ما يتكلمها الناس في البيت، بدون ما تغير المعنى:

[formal Arabic text]