TurkBench - Bias: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 250 expert-written multiple-choice items with ambiguous and disambiguated contexts testing reliance on social stereotypes (age, gender and similar themes) in Turkish. Accuracy 0-100, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiniMax-M2 | 91.4 | #327 |
| 2 | GLM-5 FP8 | 90 | #147 |
| 3 | GLM-4.6 | 88.6 | #265 |
| 4 | Qwen 3.5 122B A10B | 88.4 | #184 |
| 5 | Qwen 3.5 397B A17B | 87.4 | #154 |
| 6 | Qwen 3.5 27B | 86.2 | #192 |
| 7 | MiniMax-M2.5 | 86 | #316 |
| 8 | Qwen 3.5 35B A3B | 85.4 | #268 |
| 9 | Step 3.5 Flash | 85.2 | #281 |
| 10 | DeepSeek V3.2 Exp | 84 | #241 |
| 11 | GPT-OSS-120B | 82.8 | #346 |
| 12 | Qwen 2.5 14B Instruct | 82.8 | #654 |
| 13 | GLM-4.7 Flash | 81 | #543 |
| 14 | Gemma 3 27B (IT) | 80.13 | #528 |
| 15 | DeepSeek V3.1 | 79.6 | #275 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=turkbench-bias · How It Works · Data refreshed daily, snapshot 2026-11-05.