Toxicity Mitigation on REALTOXICITYPROMPTS
21.24ToxicityEIGENSHIFT + Targeted Subspace Intervention
Evaluation Results
| Method | Links | |
|---|---|---|
| EIGENSHIFT + Targeted Subspace InterventionBackbone=Mistral-7B-SFT2026.02 | 21.24 | |
| DETOXBackbone=GPT-2 Medium2026.02 | 24.49 | |
| DETOX + Targeted Subspace InterventionBackbone=GPT-2 Medium2026.02 | 24.61 | |
| EIGENSHIFTBackbone=Mistral-7B-SFT2026.02 | 24.64 | |
| Vanilla + Targeted Subspace InterventionBackbone=Mistral-7B-SFT2026.02 | 25.94 | |
| DETOX + Targeted Subspace InterventionBackbone=Mistral-7B-SFT2026.02 | 26.24 | |
| EIGENSHIFT + Targeted Subspace InterventionBackbone=Mistral-7B2026.02 | 28.3 | |
| DETOXBackbone=Mistral-7B-SFT2026.02 | 28.73 | |
| DETOX + Targeted Subspace InterventionBackbone=Mistral-7B2026.02 | 29.68 | |
| Vanilla + Targeted Subspace InterventionBackbone=Mistral-7B2026.02 | 31.39 | |
| DETOX + Targeted Subspace InterventionBackbone=GPT-J-6B2026.02 | 33.99 | |
| EIGENSHIFTBackbone=Mistral-7B2026.02 | 34.6 | |
| DETOXBackbone=Mistral-7B2026.02 | 34.78 | |
| Vanilla + Targeted Subspace InterventionBackbone=GPT-J-6B2026.02 | 36.21 | |
| EIGENSHIFT + Targeted Subspace InterventionBackbone=GPT-J-6B2026.02 | 38.41 | |
| VanillaBackbone=Mistral-7B-SFT2026.02 | 39.29 | |
| DETOXBackbone=GPT-J-6B2026.02 | 42.33 | |
| EIGENSHIFTBackbone=GPT-J-6B2026.02 | 44.82 | |
| VanillaBackbone=Mistral-7B2026.02 | 50.39 | |
| VanillaBackbone=GPT-J-6B2026.02 | 50.89 | |
| Vanilla + Targeted Subspace InterventionBackbone=GPT-2 Medium2026.02 | 53.93 | |
| EIGENSHIFT + Targeted Subspace InterventionBackbone=GPT-2 Medium2026.02 | 54.06 | |
| VanillaBackbone=GPT-2 Medium2026.02 | 54.25 | |
| EIGENSHIFTBackbone=GPT-2 Medium2026.02 | 54.9 |