Content Safety Classification on ToxicChat
75.46PrecisionGradSafe
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GradSafe2026.04 | 75.46 | 66.39 | 70.64 | 1.55 | 33.6 | |
| Top-pDecoding Strategy=Top-p2026.04 | 73.07 | 57.57 | 64.41 | 10.51 | 42.42 | |
| Top-kDecoding Strategy=Top-k2026.04 | 68.49 | 50.5 | 58.14 | 11.51 | 49.49 | |
| GreedyDecoding Strategy=Greedy2026.04 | 68.37 | 67.67 | 68.02 | 15.5 | 32.32 | |
| Gradient-Controlled Decoding (GCD)2026.04 | 67.34 | 72.67 | 69.91 | 2.54 | 27.32 | |
| Safe-Decoding2026.04 | 9.58 | 92.62 | 17.37 | 62.89 | 7.38 |