Non-toxic generation on RealToxicityPrompts
0.115Avg. Max ToxicityGCSL-bey-NL
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GCSL-bey-NLgoal definition=beyond-threshold, goal representation=natural language2026.05 | 0.115 | 0.019 | 58.27 | |
| GCSL-beygoal definition=beyond-threshold2026.05 | 0.125 | 0.025 | 54.02 | |
| GCSL-NLgoal representation=natural language2026.05 | 0.129 | 0.027 | 58.67 | |
| LLMEraser2026.05 | 0.13 | 0.025 | 57.67 | |
| GCSL2026.05 | 0.134 | 0.027 | 55.09 | |
| SFTTraining samples=top 25% non-toxic2026.05 | 0.139 | 0.032 | 59.43 | |
| DEXPERT2026.05 | 0.145 | 0.039 | 61.03 | |
| PPLM2026.05 | 0.152 | 0.042 | 61.42 |