Multi-ability model merging on WizardMath, Tulu-2, and WizardLM (test)
68.4Math AccuracyTask Arithmetic
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Task ArithmeticType=Ability Transfer Approach, Backbone Architecture=LLaMA 2, Parameter Scale=13B2026.01 | 68.4 | 24.6 | 31.7 | 41.6 | |
| DARE MergingType=Ability Transfer Approach, Backbone Architecture=LLaMA 2, Parameter Scale=13B2026.01 | 68.4 | 25.4 | 31.1 | 41.6 | |
| ACTType=Ability Transfer Approach, Backbone Architecture=LLaMA 2, Parameter Scale=13B, Config=transfer-only2026.01 | 66.8 | 24.3 | 34.2 | 41.8 | |
| TIES MergingType=Ability Transfer Approach, Backbone Architecture=LLaMA 2, Parameter Scale=13B2026.01 | 66.4 | 25 | 31.7 | 41 | |
| WizardMath-13BBackbone Architecture=LLaMA 2, Parameter Scale=13B, Language=English2026.01 | 63.6 | 24.1 | 12.2 | 33.3 | |
| WizardLM-13B-V1.2Backbone Architecture=LLaMA 2, Parameter Scale=13B, Language=English2026.01 | 55.2 | 23.4 | 36.6 | 38.4 | |
| Tulu-2-DPO-13BBackbone Architecture=LLaMA 2, Parameter Scale=13B, Language=English2026.01 | 34 | 28.1 | 32.3 | 31.5 |