Toxicity Classification on CivilComments
92.6Average AccuracyERM
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ERM2024.11 | 92.6 | 57.4 | — | — | |
| JTT2024.11 | 92.6 | 63.7 | — | — | |
| Robust Filtered2026.06 | 91.5 | — | 2.7 | 2.1 | |
| Robust MoE2026.06 | 91.4 | — | 2.5 | 1.8 | |
| Single Expert2026.06 | 91.3 | — | 6.9 | 6.4 | |
| Vanilla MoE2026.06 | 91.2 | — | 7.1 | 6.5 | |
| MoCaE2026.06 | 91.2 | — | 6.6 | 5.9 | |
| FGR2026.06 | 90.6 | — | 4.4 | 3.5 | |
| FGR + Robust2026.06 | 90.3 | — | 5.4 | 4.7 | |
| Bias-Conditioned Self-Influencebase_method=JTT2024.11 | 86.9 | 78.5 | — | — |