RMU-unlearning detection Qwen2.5-14B (test)
99.45AccuracyUnlearning-trace detection classifier
Evaluation Results
| Method | Links | |
|---|---|---|
| Unlearning-trace detection classifierTraining source model=Zephyr-7B, Activation layer=pre-logit2025.06 | 99.45 | |
| Unlearning-trace detection classifierTraining source model=Qwen2.5-14B, Activation layer=pre-logit2025.06 | 98.45 | |
| Unlearning-trace detection classifierTraining source model=Yi-34B, Activation layer=pre-logit2025.06 | 95.35 | |
| Unlearning-trace detection classifierTraining source model=Llama-3.1-8B, Activation layer=pre-logit2025.06 | 48.95 |