SORRY-Bench - Crimes or Torts: leaderboard
The 190 SORRY-Bench instructions that ask for help with crimes or torts (violence, intrusion, malware, fraud, weapons, terrorism and others); the number is the share the model fulfils.
Metric: Fulfillment rate on unsafe instructions (%, lower is better). Source: sorry-bench.github.io. 55 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3 Haiku (20240307) | 0 | #2314 |
| 2 | Claude 3 Opus (20240229) | 0 | #1103 |
| 3 | Claude 2.1 | 0 | #2696 |
| 4 | Claude 2 | 0 | #2427 |
| 5 | Claude 3.5 Sonnet (20240620) | 1.05 | #687 |
| 6 | Claude Instant 1.2 | 1.05 | #4019 |
| 7 | Claude 3 Sonnet (20240229) | 2.11 | #1911 |
| 8 | GPT-3.5 Turbo (0125) | 2.11 | #3500 |
| 9 | Llama 2 70B Chat | 2.11 | #6800 |
| 10 | GPT-3.5 Turbo (1106) | 2.11 | #3214 |
| 11 | O1 Mini | 2.11 | #764 |
| 12 | Gemini 1.5 Pro | 3.68 | #1049 |
| 13 | GPT-3.5 Turbo (0613) | 3.68 | #4951 |
| 14 | Llama 2 7B Chat | 3.68 | #8805 |
| 15 | O1 Preview | 3.68 | #484 |
Interactive version: theaggregate.ai/benchmark?slug=sorry-bench-crimes-or-torts · How It Works · Data refreshed daily, snapshot 2026-11-13.