Safety Classification on ToxicChat (test)
97.3AccuracyD2-TimeAttn
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| D2-TimeAttnBackbone=LLaDA-8B-Instruct, E[P]=≤0.76M2026.05 | 97.3 | 89.9 | — | — | — | |
| TimeAttnBackbone=LLaDA-1.5, E[P]=1.59M2026.05 | 97.2 | 88.6 | — | — | — | |
| D2-MLPBackbone=LLaDA-8B-Instruct, E[P]=≤0.50M2026.05 | 97.2 | 89.7 | — | — | — | |
| D2-MLPBackbone=LLaDA-1.5, E[P]=≤0.50M2026.05 | 97.2 | 89.5 | — | — | — | |
| D2-TimeAttnBackbone=LLaDA-1.5, E[P]=≤0.76M2026.05 | 97.2 | 89.7 | — | — | — | |
| MLP (Mean)Backbone=LLaDA-8B-Instruct, E[P]=1.05M2026.05 | 97 | 88 | — | — | — | |
| LSTMBackbone=LLaDA-8B-Instruct, E[P]=2.57M2026.05 | 97 | 88.9 | — | — | — | |
| D2-TimeAttnBackbone=LLaDA-8B-Base, E[P]=≤0.76M2026.05 | 97 | 88 | — | — | — | |
| MLP (Last Step)Backbone=LLaDA-8B-Instruct, E[P]=1.05M2026.05 | 96.9 | 88.1 | — | — | — | |
| D2-MLPBackbone=LLaDA-8B-Base, E[P]=≤0.50M2026.05 | 96.9 | 87.9 | — | — | — | |
| MLP (Mean)Backbone=LLaDA-1.5, E[P]=1.05M2026.05 | 96.8 | 88 | — | — | — | |
| TimeAttnBackbone=LLaDA-8B-Base, E[P]=1.59M2026.05 | 96.8 | 86.5 | — | — | — | |
| TimeAttnBackbone=LLaDA-8B-Instruct, E[P]=1.59M2026.05 | 96.8 | 88.7 | — | — | — | |
| LSTMBackbone=LLaDA-1.5, E[P]=2.57M2026.05 | 96.8 | 88 | — | — | — | |
| MLP (Last Step)Backbone=LLaDA-1.5, E[P]=1.05M2026.05 | 96.6 | 86.2 | — | — | — | |
| LSTMBackbone=LLaDA-8B-Base, E[P]=2.57M2026.05 | 96.6 | 86.7 | — | — | — | |
| MLP (MV)Backbone=LLaDA-8B-Instruct, E[P]=1.05M2026.05 | 96.3 | 87.9 | — | — | — | |
| MLP (MV)Backbone=LLaDA-1.5, E[P]=1.05M2026.05 | 96.3 | 87.8 | — | — | — | |
| MLP (Mean)Backbone=LLaDA-8B-Base, E[P]=1.05M2026.05 | 96.2 | 84.5 | — | — | — | |
| D2-TimeAttnBackbone=LLaDA-2.0-mini, E[P]=0.53M2026.05 | 96.1 | 85.2 | — | — | — | |
| D2-MLPBackbone=LLaDA-2.0-mini, E[P]=0.40M2026.05 | 96 | 84.6 | — | — | — | |
| MLP (Mean)Backbone=LLaDA-2.0-mini, E[P]=0.52M2026.05 | 95.9 | 84.6 | — | — | — | |
| LSTMBackbone=LLaDA-2.0-mini, E[P]=1.51M2026.05 | 95.9 | 82 | — | — | — | |
| MLP (Last Step)Backbone=LLaDA-8B-Base, E[P]=1.05M2026.05 | 95.8 | 82.8 | — | — | — | |
| TimeAttnBackbone=LLaDA-2.0-mini, E[P]=0.80M2026.05 | 95.8 | 80.6 | — | — | — | |
| MLP (MV)Backbone=LLaDA-8B-Base, E[P]=1.05M2026.05 | 95.5 | 84 | — | — | — | |
| MLP (Last Step)Backbone=LLaDA-2.0-mini, E[P]=0.52M2026.05 | 95.3 | 82.6 | — | — | — | |
| MLP (MV)Backbone=LLaDA-2.0-mini, E[P]=0.52M2026.05 | 95.3 | 81.9 | — | — | — | |
| LP (Last Step)Backbone=LLaDA-2.0-mini, E[P]=2×10−3M2026.05 | 93.8 | 80 | — | — | — | |
| LP (Mean)Backbone=LLaDA-2.0-mini, E[P]=2×10−3M2026.05 | 93.8 | 80.5 | — | — | — | |
| LP (MV)Backbone=LLaDA-2.0-mini, E[P]=2×10−3M2026.05 | 93.2 | 79.5 | — | — | — | |
| LP (Mean)Backbone=LLaDA-1.5, E[P]=4×10−3M2026.05 | 92.1 | 79 | — | — | — | |
| LP (Mean)Backbone=LLaDA-8B-Base, E[P]=4×10−3M2026.05 | 91.7 | 78 | — | — | — | |
| LP (Mean)Backbone=LLaDA-8B-Instruct, E[P]=4×10−3M2026.05 | 91.7 | 78.5 | — | — | — | |
| LP (MV)Backbone=LLaDA-1.5, E[P]=4×10−3M2026.05 | 91.6 | 78.2 | — | — | — | |
| LP (MV)Backbone=LLaDA-8B-Instruct, E[P]=4×10−3M2026.05 | 91.4 | 77.7 | — | — | — | |
| TPCOrder=5th order, Backbone=gemma-3-27b-it, Layer=40, Training Set=WildGuardTrain2025.09 | 91.3 | 57.66 | 44.4 | 82.24 | 7.99 | |
| LP (Last Step)Backbone=LLaDA-1.5, E[P]=4×10−3M2026.05 | 91.1 | 77.2 | — | — | — | |
| LP (MV)Backbone=LLaDA-8B-Base, E[P]=4×10−3M2026.05 | 90.7 | 76.4 | — | — | — | |
| LP (Last Step)Backbone=LLaDA-8B-Instruct, E[P]=4×10−3M2026.05 | 90.5 | 76.3 | — | — | — | |
| LP (Last Step)Backbone=LLaDA-8B-Base, E[P]=4×10−3M2026.05 | 89.8 | 74.7 | — | — | — | |
| EE-MLPOrder=5th order, Backbone=gemma-3-27b-it, Layer=40, Training Set=WildGuardTrain2025.09 | 89.42 | 52.97 | 38.95 | 82.79 | 10.07 | |
| Linear probeBackbone=gemma-3-27b-it, Layer=40, Training Set=WildGuardTrain2025.09 | 86.84 | 46.27 | 32.76 | 78.69 | 12.53 |