SALAD-Bench Base: leaderboard
SALAD-Bench — hierarchical LLM safety benchmark; base harmful-question set scored by MD-Judge safety rate.
Metric: Safety Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 2 | 99.77 | #761 |
| 2 | internlm2-chat-20B | 98.15 | #999 |
| 3 | internlm2-chat-7B | 97.7 | #1121 |
| 4 | Llama 2 13B Chat Base | 96.81 | #1422 |
| 5 | Llama 2 7B Chat | 96.51 | #1499 |
| 6 | Llama 2 70B Chat (HF) | 96.21 | #1264 |
| 7 | gemma-2B (IT) | 95.9 | #1559 |
| 8 | Qwen 1.5 4B Chat | 95.51 | #1467 |
| 9 | Qwen 1.5 14B Chat | 95.37 | #977 |
| 10 | Qwen-14B-Chat | 95.35 | #1169 |
| 11 | Qwen-72B-Chat | 94.4 | #843 |
| 12 | gemma-7B (IT) | 94.08 | #1496 |
| 13 | Qwen 1.5 72B Chat | 93.55 | #776 |
| 14 | GPT-4 Turbo | 93.49 | #404 |
| 15 | Qwen 1.5 7B Chat | 93 | #1324 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=salad-bench-base · How It Works · Data refreshed daily, snapshot 2026-10-20.