HELM SeaHELM - Thai Toxicity Tweets: leaderboard
HELM SeaHELM Southeast-Asian-languages evaluation: Thai-language toxic-tweet detection.
Metric: Macro F1 score. Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet (20250219) | 71.23 | #201 |
| 2 | GPT-4o (2024-11-20) | 70.52 | #381 |
| 3 | Qwen 2.5 72B Instruct | 68.48 | #454 |
| 4 | Qwen 2 72B Instruct | 68.01 | #576 |
| 5 | Gemini 2.0 Flash Lite | 66.38 | #450 |
| 6 | Llama 4 Maverick Instruct FP8 | 65.93 | #414 |
| 7 | Llama 3.1 405B Instruct | 64.92 | #457 |
| 8 | GPT-4o Mini (2024-07-18) | 64.11 | #558 |
| 9 | Llama 3.2 90B Vision Instruct | 63.53 | #622 |
| 10 | Llama 4 Scout Instruct | 61.09 | #539 |
| 11 | Llama 3.3 70B Instruct | 60.59 | #534 |
| 12 | Gemini 2.0 Flash | 59.39 | #342 |
| 13 | Llama 3.1 70B Instruct | 59.31 | #561 |
| 14 | Llama 3.1 8B CPT SEA LIONv3 Instruct | 58.26 | #810 |
| 15 | DeepSeek V3 | 55.11 | #321 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-seahelm-thai-toxicity-tweets · How It Works · Data refreshed daily, snapshot 2026-10-20.