HELM Arabic - Aratrust - Offensive: leaderboard
HELM Arabic (release v2.2.0) AraTrust Offensive category: human-written Arabic three-option multiple-choice questions probing trustworthiness on offensive language; 69 questions, exact match. Percent correct, 0-100, higher is better; guessing scores 33.3.
Metric: Exact Match (%). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 41 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Maverick Instruct FP8 | 95.65 | #424 |
| 2 | jais-family-30B-16k-chat | 95.65 | #758 |
| 3 | Mistral Large 3 | 94.2 | #410 |
| 4 | Gemini 2.5 Flash Lite (Non-reasoning) | 94.2 | #432 (Gemini 2.5 Flash Lite) |
| 5 | Gemini 2.5 Flash (Non-reasoning) | 92.75 | #255 (Gemini 2.5 Flash) |
| 6 | jais-adapted-7B-chat | 92.75 | #1160 |
| 7 | Claude Haiku 4.5 (20251001) | 91.3 | #264 |
| 8 | GPT-5.1 (2025-11-13) | 89.86 | #107 |
| 9 | Llama 3.3 70B Instruct | 88.41 | #541 |
| 10 | Gemma 4 31B (IT) | 88.41 | #178 |
| 11 | DeepSeek V3.1 | 88.41 | #275 |
| 12 | GPT-4.1 Nano | 86.96 | #730 |
| 13 | Qwen 3.5 397B A17B | 86.96 | #154 |
| 14 | Mistral Large 2 (Nov) Instruct (2411) | 86.96 | #480 |
| 15 | AceGPT-v2-32B-Chat | 86.96 | #518 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-arabic-aratrust-offensive · How It Works · Data refreshed daily, snapshot 2026-11-05.