ThaiSafetyBench - Discrimination, Exclusion, Toxicity, Hateful, Offensive: leaderboard
ThaiSafetyBench (SCB DataX and SCB 10X, 2026) sends 1,954 harmful prompts written in Thai to each model and judges the responses with GPT-4.1 and Gemini 2.5 Pro; the attack success rate is averaged over both judges. This board covers the prompts in its discrimination, exclusion, toxicity, hateful and offensive content harm category. Metric is attack success rate (%), lower is better (safer).
Metric: Attack Success Rate (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 2.19 | #105 |
| 2 | Claude Sonnet 4.5 | 3.49 | #148 |
| 3 | Llama-SEA-LION-v3-70B-IT | 5.48 | #407 |
| 4 | Qwen 2.5 72B Instruct | 6.27 | #469 |
| 5 | Qwen 2.5 7B Instruct | 6.97 | #863 |
| 6 | GPT-4o | 7.87 | #357 |
| 7 | Llama-SEA-LION-v3-8B-IT | 8.57 | #800 |
| 8 | Llama 3.3 70B Instruct | 8.67 | #541 |
| 9 | Gemma 3 12B (IT) | 8.86 | #672 |
| 10 | SeaLLMs-v3-7B-Chat | 8.96 | #897 |
| 11 | Llama 3.1 70B Instruct | 15.64 | #567 |
| 12 | Gemma 3 4B (IT) | 18.55 | #997 |
| 13 | Llama 3.2 3B Instruct | 24.32 | #1327 |
| 14 | Llama 3.1 8B Instruct | 25.5 | #1017 |
| 15 | Llama 3.2 1B Instruct | 36.59 | #1518 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=thaisafetybench-discrimination-exclusion-toxicity-hateful-offensive · How It Works · Data refreshed daily, snapshot 2026-11-05.