ArabCulture-Dialogue - Dialect Steering Quality: leaderboard

ArabCulture-Dialogue (MBZUAI, 2026) tests dialect steering: given a dialogue context and a Modern Standard Arabic utterance, the model writes the next turn in the dialect of a named Arab country. A GPT-5 judge rates each continuation from 1 to 5, rescaled to 0-100; higher is better.

Metric: Continuation quality (0-100): GPT-5 judge rating on 1-5 rescaled as (s-1)/4 and reported times 100, zero-shot dialect steering: given a dialogue context and an MSA utterance, the model writes one continuation in the target country's dialect, over the dialogues of ArabCulture-Dialogue (13 Arab countries, parallel Modern Standard Arabic and country-dialect versions of 3,471 culturally grounded multi-turn dialogues built from ArabCulture); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-595.6#178
2Gemini 2.5 Pro92.4#276
3ALLaM-7B-Instruct-preview81.68#2001
4Jais-2-8B-Chat73#2496
5c4ai-command-r7B-arabic-02-202571.6#2135
6Fanar-1-9B62.1#1913
7Hala-9B61.2#1162
8Gemma 2 9B (IT)60.4#2369
9jais-adapted-7B-chat59.9#5831
10SILMA-9B-Instruct-v1.059.1#3847
11Llama 3.1 8B Instruct51.6#4600
12Qwen 3 8B43#1873

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=arabculture-dialogue-dialect-steering-quality · How It Works · Data refreshed daily, snapshot 2026-11-21.