Input Moderation on ToxicChat (test)
82.8F1 ScoreQwen3Guard-4B-Gen-loose
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen3Guard-4B-Gen-looseStreaming=no2026.04 | 82.8 | — | |
| Qwen3Guard-8B-Gen-looseStreaming=no2026.04 | 82.8 | — | |
| Qwen3Guard-Stream-4B-looseStreaming=yes2026.04 | 81.7 | — | |
| Qwen3Guard-Stream-8B-looseStreaming=yes2026.04 | 80.1 | — | |
| Qwen3Guard-0.6B-Gen-looseStreaming=no2026.04 | 77.7 | — | |
| Llama3-StreamGuard-3BStreaming=yes2026.04 | 77.7 | — | |
| Llama3-StreamGuard-8BStreaming=yes2026.04 | 77.4 | — | |
| NemoGuard-8BStreaming=no2026.04 | 75.6 | — | |
| Qwen3Guard-Stream-0.6B-looseStreaming=yes2026.04 | 75.5 | — | |
| Qwen3Guard-Stream-8B-strictStreaming=yes2026.04 | 75.3 | — | |
| D2-MLPBackbone=LLaDA-8B-Instruct, E[P]=≤0.25M2026.05 | 75 | 90.4 | |
| Llama3-StreamGuard-1BStreaming=yes2026.04 | 74.9 | — | |
| D2-TimeAttnBackbone=LLaDA-8B-Instruct, E[P]=≤0.85M2026.05 | 74.4 | 89.9 | |
| TimeAttnBackbone=LLaDA-8B-Instruct, E[P]=1.59M2026.05 | 73.1 | 88.9 | |
| Qwen3Guard-Stream-4B-strictStreaming=yes2026.04 | 73 | — | |
| ShieldGemma-27BStreaming=no2026.04 | 72.9 | — | |
| Qwen3Guard-Stream-0.6B-strictStreaming=yes2026.04 | 72 | — | |
| LSTMBackbone=LLaDA-8B-Instruct, E[P]=2.57M2026.05 | 72 | 88.6 | |
| PolyGuard-Qwen-7BStreaming=no2026.04 | 71.5 | — | |
| D2-TimeAttnBackbone=LLaDA-8B-Base, E[P]=≤0.85M2026.05 | 71 | 87.5 | |
| WildGuard-7BStreaming=no2026.04 | 70.8 | — | |
| MLP (Mean)Backbone=LLaDA-8B-Instruct, E[P]=1.05M2026.05 | 70.8 | 86.8 | |
| TimeAttnBackbone=LLaDA-8B-Base, E[P]=1.59M2026.05 | 70.8 | 87.1 | |
| D2-MLPBackbone=LLaDA-8B-Base, E[P]=≤0.25M2026.05 | 70.7 | 87 | |
| MLP (Last Step)Backbone=LLaDA-8B-Instruct, E[P]=1.05M2026.05 | 70.6 | 87.3 | |
| MLP (MV)Backbone=LLaDA-8B-Instruct, E[P]=1.05M2026.05 | 70.1 | 86.3 | |
| Qwen3Guard-4B-Gen-strictStreaming=no2026.04 | 69.5 | — | |
| ShieldGemma-9BStreaming=no2026.04 | 69.4 | — | |
| LP (Last Step)Backbone=LLaDA-8B-Instruct, E[P]=4×10−3M2026.05 | 69 | 86.1 | |
| Qwen3Guard-8B-Gen-strictStreaming=no2026.04 | 68.9 | — | |
| LP (Mean)Backbone=LLaDA-8B-Instruct, E[P]=4×10−3M2026.05 | 68.9 | 85.3 | |
| LP (Mean)Backbone=LLaDA-8B-Base, E[P]=4×10−3M2026.05 | 68.1 | 85.4 | |
| LP (MV)Backbone=LLaDA-8B-Instruct, E[P]=4×10−3M2026.05 | 67.6 | 84.5 | |
| MLP (Mean)Backbone=LLaDA-8B-Base, E[P]=1.05M2026.05 | 67.5 | 84.6 | |
| LP (MV)Backbone=LLaDA-8B-Base, E[P]=4×10−3M2026.05 | 66.9 | 84.5 | |
| MLP (Last Step)Backbone=LLaDA-8B-Base, E[P]=1.05M2026.05 | 66.3 | 84 | |
| LSTMBackbone=LLaDA-8B-Base, E[P]=2.57M2026.05 | 66.1 | 83.1 | |
| MLP (MV)Backbone=LLaDA-8B-Base, E[P]=1.05M2026.05 | 65.3 | 82.6 | |
| Qwen3Guard-0.6B-Gen-strictStreaming=no2026.04 | 65.1 | — | |
| LP (Last Step)Backbone=LLaDA-8B-Base, E[P]=4×10−3M2026.05 | 64.8 | 83.2 | |
| LlamaGuard3-8BStreaming=no2026.04 | 53.8 | — | |
| LlamaGuard4-12BStreaming=no2026.04 | 51.3 | — |