ML-Bench (Multilingual Safety) - Seed Query F1: leaderboard
Metric: Unsafe-query F1 (%; safe/unsafe classification of policy-grounded seed queries in 14 languages, guardrail given only the input text). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 12 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | ML-Guard-7B | 95 |
| 2 | ML-Guard-1.5B | 85 |
| 3 | PolyGuard-Qwen | 72 |
| 4 | Llama-3.1-Nemotron-Safety-Guard-8B-v3 | 62 |
| 5 | Qwen3Guard-Gen-8B | 61 |
| 6 | Qwen3Guard-Gen-4B | 59 |
| 7 | Qwen3Guard-Gen-0.6B | 57 |
| 8 | Llama-Guard-3-8B | 49 |
| 9 | Llama-Guard-3-1B | 48 |
| 10 | DuoGuard-1.5B-transfer | 46 |
Interactive version: theaggregate.ai/benchmark?slug=ml-bench-multilingual-safety-seed-query-f1 · How It Works · Data refreshed daily, snapshot 2026-09-26.