Safety-Flag: leaderboard
Safety-Flag (Hu, 2026) recasts seven safety datasets (BeaverTails, XSTest, Ethics, WildGuard, Aegis, ToxiChat and ToxiGen) into one balanced flag or do-not-flag moderation task on identical items with a standardized prompt. Macro-F1 in percent averaged over the seven sources; higher is better.
Metric: Macro-F1 (%; mean over the seven source benchmarks of the macro-F1 of flag / do-not-flag verdicts on 198-200 class-balanced items from each of BeaverTails, XSTest, Ethics, WildGuard, Aegis, ToxiChat and ToxiGen, recast as one flag / do-not-flag decision with harmonised source labels; one standardized two-option prompt with a JSON verdict and confidence; greedy decoding). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 85.3 | #389 |
| 2 | GPT-4.1 Mini | 84.3 | #775 |
| 3 | Gemma 2 9B (IT) | 82.4 | #2275 |
| 4 | DeepSeek R1 Distill Llama 8B | 81.3 | #6773 |
| 5 | Qwen 2.5 7B Instruct | 79.3 | #2583 |
| 6 | Mistral 7B Instruct (v0.3) | 73.3 | #6698 |
| 7 | OLMo-2-1124-7B-Instruct | 63.9 | #7377 |
| 8 | Llama 3.1 8B Instruct | 44.3 | #4277 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=safety-flag · How It Works · Data refreshed daily, snapshot 2026-11-12.