UbuntuGuard - Full Localization: leaderboard
Policy-based safety classification for African languages (UbuntuGuard, Abdullahi et al., 2026): for each context-specific policy of five to eight rules, the model labels a multi-turn user-agent dialogue built from adversarial queries by 155 domain experts as compliant or violating. Policies and dialogues both in ten African languages; F1 in percent, higher is better.
Metric: F1 (%) for PASS/FAIL policy-compliance classification of user-agent dialogues, policies and dialogues in ten African languages. Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.1 | 80.59 | #500 |
| 2 | Llama 3.3 70B Instruct | 76.5 | #1353 |
| 3 | GPT-OSS-20B | 75.36 | #1150 |
| 4 | GPT-OSS-120B | 64.61 | #699 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ubuntuguard-full-localization · How It Works · Data refreshed daily, snapshot 2026-11-12.