Toxicity Mitigation on RealToxicityPrompts (RTP)
0.53CLS Tox RateCHaRS
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CHaRSBackbone=Gemma2-2B2026.02 | 0.53 | 3.63 | |
| CHaRS-PCTBackbone=Gemma2-2B2026.02 | 0.57 | 3.47 | |
| Linear-AcTBackbone=Gemma2-2B2026.02 | 0.67 | 4.6 | |
| CHaRSBackbone=Qwen2.5-7B2026.02 | 1.03 | 4.97 | |
| CHaRS-PCTBackbone=Qwen2.5-7B2026.02 | 1.07 | 4.37 | |
| CHaRS-PCTBackbone=Llama3-8B2026.02 | 1.17 | 4.47 | |
| CHaRSBackbone=Llama3-8B2026.02 | 1.23 | 4.8 | |
| Linear-AcTBackbone=Qwen2.5-7B2026.02 | 1.8 | 6.1 | |
| Linear-AcTBackbone=Llama3-8B2026.02 | 1.93 | 7.73 | |
| OriginalBackbone=Gemma2-2B2026.02 | 2.93 | 10.77 | |
| OriginalBackbone=Qwen2.5-7B2026.02 | 3.43 | 10.27 | |
| OriginalBackbone=Llama3-8B2026.02 | 4.8 | 14.6 |