COMPL-AI - Toxic Completion Avoidance: leaderboard
The model continues RealToxicityPrompts openings; the score is one minus the classifier-rated toxicity of the continuations. It measures unprompted toxic output rather than refusal of a harmful request.
Metric: Non-toxicity of completions (%, higher is better). Source: huggingface.co. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral 7B Instruct (v0.3) | 97.04 | #6653 |
| 2 | Qwen 1.5 72B Chat | 96.96 | #2126 |
| 3 | GPT-4 Turbo (Preview) | 96.93 | #1083 |
| 4 | Mistral 7B Instruct (v0.2) | 96.47 | #7007 |
| 5 | Llama 2 7B Chat | 96.47 | #8805 |
| 6 | Llama 2 13B Chat | 96.37 | #8186 |
| 7 | Mixtral 8x7B Instruct | 95.81 | #4444 |
| 8 | Gemini 1.5 Flash (001) | 94.17 | #1784 |
| 9 | Llama 2 70B Chat | 94.09 | #6800 |
| 10 | GPT-3.5 Turbo (0125) | 93.88 | #3500 |
| 11 | Bielik-11B-v2.3-Instruct | 93.65 | #1791 |
| 12 | Yi Chat 34B | 92.19 | #3544 |
| 13 | Mistral-7B-v0.3 | 87.92 | #7191 |
| 14 | Gemma 2 9B | 84.15 | #2802 |
Interactive version: theaggregate.ai/benchmark?slug=compl-ai-toxic-completion-avoidance · How It Works · Data refreshed daily, snapshot 2026-11-13.