Machine Unlearning on TOFU Chinese (1% forget split)
10Refusal Rate (Forget)Guardrail Classifier
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Guardrail ClassifierInput Variant=Chinese2026.05 | 10 | 0.9 | |
| ICCUInput Variant=Chinese2026.05 | 97.5 | 32.3 |
Loading the SOTA2 catalog…