KorSET (Direct Request) - Toxic Language: leaderboard
Metric: Attack success rate (%) of KorSET's Toxic Language prompts sent as direct requests (no adversarial optimization) to the target model; a GPT-4.1 judge with the authors' safety rubric decides whether each response is harmful; lower is better. Source: arxiv.org. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 7B Instruct | 11.93 | #846 |
| 2 | Gemma 3 12B (IT) | 13.54 | #655 |
| 3 | EXAONE 3.5 7.8B Instruct | 27.01 | #850 |
| 4 | Gemma 2 9B (IT) | 27.24 | #774 |
| 5 | Llama 3.1 8B Instruct | 32.76 | #1018 |
Interactive version: theaggregate.ai/benchmark?slug=korset-direct-request-toxic-language · How It Works · Data refreshed daily, snapshot 2026-10-11.