TurkBench - Toxicity: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 250 expert-written harmful or manipulative Turkish requests (for example workplace harassment); responses are graded by an LLM judge (0-100, higher is better), where higher means fewer harmful, offensive or inappropriate outputs. Scores cluster above 90.
Metric: LLM-as-a-judge score (0-100). Source: huggingface.co. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 0.6B | 100 | #1448 |
| 2 | GPT-OSS-120B | 99.86 | #346 |
| 3 | Qwen 3.5 122B A10B | 99.53 | #184 |
| 4 | Qwen 3.5 35B A3B | 99.18 | #268 |
| 5 | DeepSeek V3.2 Exp | 99.14 | #241 |
| 6 | Qwen 2.5 14B Instruct | 99.02 | #654 |
| 7 | Qwen 3.5 27B | 98.87 | #192 |
| 8 | MiniMax-M2 | 98.8 | #327 |
| 9 | Qwen 3.5 397B A17B | 98.74 | #154 |
| 10 | DeepSeek V3.1 | 98.66 | #275 |
| 11 | Llama 3.1 8B Instruct | 98.58 | #1017 |
| 12 | DeepSeek R1 0528 Qwen3 8B | 98.52 | #765 |
| 13 | aya-expanse-8B | 98.4 | #914 |
| 14 | Phi-4 Mini Instruct | 98.4 | #1232 |
| 15 | Qwen 3 1.7B | 98.13 | #1175 |
Interactive version: theaggregate.ai/benchmark?slug=turkbench-toxicity · How It Works · Data refreshed daily, snapshot 2026-11-05.