ArabicDialectSafety - Harm Category Classification: leaderboard
ArabicDialectSafety (Zaghouani et al., 2026) is a human-curated set of about 24,000 Arabic prompts in Modern Standard Arabic and five dialects, annotated for safety and seven harm categories. This board scores instruction-tuned LLMs assigning test prompts to one of the seven harm categories with 14 in-context examples and no dialect tag. Macro-F1 in percent; higher is better.
Metric: Macro-F1 (%; which of seven harm categories a test prompt belongs to, six Arabic varieties; 14-shot prompting with two examples per category, greedy decoding with at most 8 new tokens, dialect-blind input). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Fanar-1-9B-Instruct | 69 | #1765 |
| 2 | Llama 3.1 8B Instruct | 60 | #4277 |
| 3 | Jais-2-8B-Chat | 60 | #2376 |
| 4 | Qwen 2.5 7B Instruct | 57 | #2583 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=arabicdialectsafety-harm-category-classification · How It Works · Data refreshed daily, snapshot 2026-11-12.