Machine Unlearning on TOFU Llama 3.1 8B (5%) benchmark leaderboard · SOTA2 Research

Machine Unlearning on TOFU Llama 3.1 8B (5%)

0.914FQT3-Unlearning

Evaluation Results

Method	Links
T3-Unlearning2026.02		0.914	0.9	0.612	41.5	0.983	0.461	54.5	0.874	0.688	51.1	0.853	0.648
NPO2026.02		0.793	0.713	0.631	77.9	0.59	0.507	54.7	0.846	0.719	50.4	0.751	0.638
SimNPO2026.02		0.745	0.786	0.653	83	0.668	0.508	54.9	0.868	0.72	50.5	0.856	0.634
IdkDPO2026.02		0.535	0.669	0.545	67.8	0.534	0.481	42.9	0.833	0.585	40.2	0.708	0.518
RMU2026.02		0.532	0	0	0	0.038	0.275	18.5	0.005	0.449	20.2	0	0.516
SatImp2026.02		0.445	0.419	0.412	40.7	0.436	0.413	36.8	0.612	0.489	35.1	0.313	0.465
WGA2026.02		0.252	0.424	0.418	40.5	0.436	0.408	37.4	0.617	0.502	37.1	0.327	0.478
ULD2026.02		0.169	0.944	0.644	98.8	0.974	0.534	49.8	0.91	0.641	41.3	0.95	0.518
UNDIAL2026.02		0.016	0.653	0.567	60.1	0.47	0.46	47.9	0.832	0.645	47	0.795	0.596
GradDiff2026.02		0.003	0.281	0.363	53.6	0.529	0.452	35.9	0.499	0.477	35.9	0.153	0.48
Original Model2026.02		0	0.948	0.637	99.1	0.995	0.531	47.9	0.905	0.625	40.9	0.949	0.514
GradAscent2026.02		0	0.875	0.576	92.5	0.852	0.525	42	0.888	0.543	35.2	0.885	0.463