TSHA - Open-Ended Hazard QA: leaderboard
Metric: Open-ended QA overall (%): GPT-4o rubric score, 0.7 hazard accuracy + 0.2 conciseness + 0.1 coherence, on the 1,707-item TSHA test set (existing indoor datasets, new photos, internet and AIGC images, Sora videos, Hunyuan panoramas); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 78.8 | #194 |
| 2 | Gemini 2.5 Pro | 78.3 | #145 |
| 3 | Gemini 2.5 Flash | 77.1 | #237 |
| 4 | Qwen 2.5 VL 32B Instruct | 75.7 | #443 |
| 5 | Claude 3.7 Sonnet | 74 | #241 |
| 6 | Qwen 2.5 VL 7B Instruct | 71.1 | #643 |
| 7 | InternVL3-8B | 68.9 | #606 |
| 8 | Gemma 3 27B (IT) | 37.7 | #509 |
| 9 | Gemma 3 4B (IT) | 36.7 | #971 |
| 10 | Gemma 3 12B (IT) | 36.4 | #655 |
| 11 | Mistral Small 3.1 | 19.4 | #600 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=tsha-open-ended-hazard-qa · How It Works · Data refreshed daily, snapshot 2026-10-11.