Safety on XSTest (Normalized performance)
1.245Normalized PerformanceMETIS
Evaluation Results
| Method | Links | |
|---|---|---|
| METISCategory=Ours, Backbone=Llama-3.2-3B2026.06 | 1.245 | |
| TSV-MCategory=State-of-the-Art (+Many-shot), Backbone=Llama-3.2-3B2026.06 | 1.184 | |
| ConsensusTACategory=Conventional Method (+Many-shot), Backbone=Llama-3.2-3B2026.06 | 1.163 | |
| TIESCategory=Conventional Method (+Many-shot), Backbone=Llama-3.2-3B2026.06 | 1.143 | |
| TULU 3Category=Data-Mixing, Backbone=Llama-3.2-3B2026.06 | 1.102 | |
| DARECategory=Conventional Method (+Many-shot), Backbone=Llama-3.2-3B2026.06 | 0.98 | |
| TA + Subspace BoostingCategory=State-of-the-Art (+Many-shot), Backbone=Llama-3.2-3B2026.06 | 0.959 | |
| Task ArithmeticCategory=Conventional Method (+Many-shot), Backbone=Llama-3.2-3B2026.06 | 0.878 | |
| ColD FusionCategory=Iterative / FL, Backbone=Llama-3.2-3B2026.06 | 0.857 | |
| Iso-CCategory=State-of-the-Art (+Many-shot), Backbone=Llama-3.2-3B2026.06 | 0.837 | |
| q-FedAvgCategory=Iterative / FL, Backbone=Llama-3.2-3B2026.06 | 0.837 | |
| FedMergeCategory=Iterative / FL, Backbone=Llama-3.2-3B2026.06 | 0.796 | |
| Iso-CTSCategory=State-of-the-Art (+Many-shot), Backbone=Llama-3.2-3B2026.06 | 0.653 |