CoCoNot - Safety Concerns: leaderboard
The model receives CoCoNot requests with safety concerns (harmful, offensive, privacy-violating, copyright or misinformation requests), and the score is the share it complies with, judged by GPT-3.5-turbo against category rules. It is the should-refuse side matched to the CoCoNot contrast set, on the same models.
Metric: Compliance rate on requests with safety concerns (%, lower is better). Source: arxiv.org. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 Preview (1106) | 2 | #953 |
| 2 | Claude 3 Sonnet | 6.3 | #2309 |
| 3 | Llama 2 7B Chat | 7.1 | #8805 |
| 4 | GPT-3.5 Turbo (0125) | 8.1 | #3500 |
| 5 | tulu-2-dpo-70B | 8.1 | #3944 |
| 6 | Llama 2 13B Chat | 9.1 | #8186 |
| 7 | Llama 2 70B Chat | 10.1 | #6800 |
| 8 | GPT-4 (0613) | 11.4 | #1214 |
| 9 | tulu-2-dpo-7B | 11.7 | #7492 |
| 10 | tulu-2-dpo-13B | 12.2 | #6621 |
| 11 | GPT-4o (2024-05-13) | 12.7 | #735 |
| 12 | Llama 3 8B Instruct | 13.2 | #4612 |
| 13 | tulu-2-13B | 13.7 | #7711 |
| 14 | gemma-7B (IT) | 13.9 | #8825 |
| 15 | tulu-2-7B | 17 | #8164 |
Interactive version: theaggregate.ai/benchmark?slug=coconot-safety-concerns · How It Works · Data refreshed daily, snapshot 2026-11-13.