Detoxification on RealToxicityPrompts
0.27Avg Max ToxicitySGEAT + DEXPERTS
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SGEAT + DEXPERTSBackbone=1.3B LM, Category=Combined2022.02 | 0.27 | 0.14 | — | |
| DEXPERTSexpert_size=large2021.05 | 0.293 | 0.111 | — | |
| DEXPERTSBackbone=1.3B LM, Category=Decoding-Time, Decoding Speed=3x slow2022.02 | 0.31 | 0.18 | — | |
| SGEAT + Rejection SamplingBackbone=1.3B LM, Category=Combined2022.02 | 0.33 | 0.21 | — | |
| SGEAT (augmented)Backbone=1.3B LM, Category=Domain-Adaptive Training, Prompt Design=augmented2022.02 | 0.43 | 0.37 | — | |
| SGEAT (standard)Backbone=1.3B LM, Category=Domain-Adaptive Training, Prompt Design=standard2022.02 | 0.44 | 0.38 | — | |
| Rejection SamplingBackbone=1.3B LM, Category=Decoding-Time, Decoding Speed=4x slow2022.02 | 0.45 | 0.39 | — | |
| SGEAT + DEXPERTSBackbone=1.3B LM, Category=Combined2022.02 | 0.45 | 0.4 | — | |
| DAPT (nontoxic)Backbone=1.3B LM, Category=Domain-Adaptive Training2022.02 | 0.47 | 0.43 | — | |
| SGEAT (heuristic)Backbone=1.3B LM, Category=Domain-Adaptive Training, Prompt Design=heuristic2022.02 | 0.47 | 0.43 | — | |
| DEXPERTSBackbone=1.3B LM, Category=Decoding-Time, Decoding Speed=3x slow2022.02 | 0.5 | 0.47 | — | |
| GPT-32021.05 | 0.525 | 0.515 | — | |
| Word BanningBackbone=1.3B LM, Category=Decoding-Time2022.02 | 0.54 | 0.56 | — | |
| SGEAT + Rejection SamplingBackbone=1.3B LM, Category=Combined2022.02 | 0.56 | 0.58 | — | |
| Jigsaw (nontoxic)Backbone=1.3B LM, Category=Domain-Adaptive Training2022.02 | 0.58 | 0.61 | — | |
| SGEAT (standard)Backbone=1.3B LM, Category=Domain-Adaptive Training, Prompt Design=standard2022.02 | 0.67 | 0.75 | — | |
| SGEAT (augmented)Backbone=1.3B LM, Category=Domain-Adaptive Training, Prompt Design=augmented2022.02 | 0.68 | 0.77 | — | |
| Rejection SamplingBackbone=1.3B LM, Category=Decoding-Time, Decoding Speed=4x slow2022.02 | 0.68 | 0.78 | — | |
| DAPT (nontoxic)Backbone=1.3B LM, Category=Domain-Adaptive Training2022.02 | 0.69 | 0.79 | — | |
| Word BanningBackbone=1.3B LM, Category=Decoding-Time2022.02 | 0.72 | 0.86 | — | |
| SGEAT (heuristic)Backbone=1.3B LM, Category=Domain-Adaptive Training, Prompt Design=heuristic2022.02 | 0.73 | 0.85 | — | |
| Jigsaw (nontoxic)Backbone=1.3B LM, Category=Domain-Adaptive Training2022.02 | 0.77 | 0.9 | — | |
| CAAModel=Gemma-2-9b-pt2025.05 | — | — | 73.98 | |
| CAAModel=Gemma-2-9b-it2025.05 | — | — | 98.75 | |
| CAAModel=Llama-3.1-8B2025.05 | — | — | 72.81 | |
| Prompt_autoModel=Gemma-2-9b-pt2025.05 | — | — | 57.63 | |
| Prompt_autoModel=Gemma-2-9b-it2025.05 | — | — | 98.92 | |
| Prompt_autoModel=Llama-3.1-8B2025.05 | — | — | 60.64 | |
| Prompt_handModel=Gemma-2-9b-pt2025.05 | — | — | 53.96 | |
| Prompt_handModel=Gemma-2-9b-it2025.05 | — | — | 98.42 | |
| Prompt_handModel=Llama-3.1-8B2025.05 | — | — | 57.3 | |
| SAE_AXBENCHModel=Gemma-2-9b-pt2025.05 | — | — | 75.15 | |
| SAE_AXBENCHModel=Gemma-2-9b-it2025.05 | — | — | 98.42 | |
| SAE_AXBENCHModel=Llama-3.1-8B2025.05 | — | — | 71.98 | |
| STAModel=Gemma-2-9b-pt2025.05 | — | — | 76.98 | |
| STAModel=Gemma-2-9b-it2025.05 | — | — | 99.33 | |
| STAModel=Llama-3.1-8B2025.05 | — | — | 73.64 | |
| VanillaModel=Gemma-2-9b-pt2025.05 | — | — | 57.63 | |
| VanillaModel=Gemma-2-9b-it2025.05 | — | — | 97.41 | |
| VanillaModel=Llama-3.1-8B2025.05 | — | — | 58.38 |