ThaiSafetyBench - Thai Socio-Cultural Harms: leaderboard
Metric: Attack Success Rate (%). Source: huggingface.co. 24 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 8.97 |
| 2 | SeaLLMs-v3-7B-Chat | 16.98 |
| 3 | Claude Sonnet 4.5 | 19.47 |
| 4 | Qwen 2.5 72B Instruct | 22.9 |
| 5 | openthaigpt1.5-72B-instruct | 26.15 |
| 6 | GPT-4o | 27.48 |
| 7 | Qwen 2.5 7B Instruct | 33.4 |
| 8 | openthaigpt1.5-7B-instruct | 36.07 |
| 9 | Gemma 3 12B (IT) | 38.55 |
| 10 | llama3.1-typhoon2-70B-instruct | 43.13 |
| 11 | Llama 3.3 70B Instruct | 44.66 |
| 12 | Gemma 3 4B (IT) | 47.14 |
| 13 | Llama 3.1 70B Instruct | 54.39 |
| 14 | Llama 3.2 3B Instruct | 57.37 |
| 15 | llama3.2-typhoon2-3B-instruct | 58.02 |
Interactive version: theaggregate.ai/benchmark?slug=thaisafetybench-thai-socio-cultural-harms · How It Works · Data refreshed daily, snapshot 2026-09-19.