Text Detoxification on Our Dataset
100STAhuman-baseline
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| human-baseline2026.06 | 100 | 95.2 | 89.6 | 85.4 | |
| Tatoxa2026.06 | 97 | 85.8 | 80.7 | 68 | |
| mT0+Gemini Pro v2.52026.06 | 95 | 81.4 | 80.5 | 64.2 | |
| GPT-5 Chat2026.06 | 93.3 | 81.2 | 83.6 | 64.2 | |
| Claude Opus 4.62026.06 | 91.6 | 86 | 82.5 | 66 | |
| Gemini Pro v2.52026.06 | 89.7 | 85.5 | 83.8 | 65.3 | |
| mT0+vocab deletion2026.06 | 86.3 | 84.8 | 80.2 | 59.8 | |
| Vocab deletion2026.06 | 80 | 89.5 | 81.5 | 58.5 | |
| DeepSeek V3.22026.06 | 78 | 90.5 | 82.4 | 58.9 | |
| mT0prompt=Russian prompt2026.06 | 76.4 | 84.8 | 81.8 | 53.6 | |
| mT0prompt=Tatar prompt2026.06 | 75.8 | 85.3 | 81.8 | 53.5 | |
| mT0prompt=English prompt2026.06 | 74.2 | 87 | 82.3 | 53.7 |