Olmo 3 Safety Evaluation - WildGuard-Test: leaderboard
The model answers the 749 adversarial prompts of WildGuard-Test (synthetic and in-the-wild, harmful and benign); the score is the share of responses WildGuard labels safe. Unlike WildJailbreak, the prompts were collected from real user interactions as well as synthetic attacks.
Metric: Safe responses (%, 749 adversarial prompts, WildGuard safety label, higher is better). Source: arxiv.org. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | OLMo 3 7B Instruct | 99.6 | #6860 |
| 2 | OLMo-2-1124-7B-Instruct | 99.2 | #7343 |
| 3 | Olmo-3.1-32B-Instruct-SFT | 98.9 | #3407 |
| 4 | Olmo-3.1-32B-Instruct-DPO | 98.3 | #3174 |
| 5 | OLMo-2-0325-32B-Instruct | 98.3 | #2570 |
| 6 | OLMo 3.1 32B Instruct | 97.8 | #2384 |
| 7 | OLMo-3.1 32B Think | 96.9 | #1443 |
| 8 | Apertus-70B-Instruct-2509 | 95.4 | #2639 |
| 9 | Granite 3.3 8B Instruct | 93.8 | #3615 |
| 10 | OLMo 3 7B (Thinking) | 93.8 | #3899 (OLMo 3 7B) |
| 11 | Qwen 3 VL 32B Instruct | 93.5 | #585 |
| 12 | Qwen 3 VL 8B (Thinking) | 93 | #887 |
| 13 | Qwen 3 VL 32B (Thinking) | 92.9 | #559 |
| 14 | Gemma 2 27B (IT) | 92 | #1782 |
| 15 | Qwen 2.5 32B Instruct | 91.7 | #1215 |
Interactive version: theaggregate.ai/benchmark?slug=olmo-3-safety-evaluation-wildguard-test · How It Works · Data refreshed daily, snapshot 2026-11-13.