Task addition on Task Arithmetic Benchmark (test)
88.5Avg Absolute AccuracyLinear. FT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Linear. FTBackbone=ViT-L/14, Model formulation=f_lin(·; θ₀ + τ_lin)2023.05 | 88.5 | 93.5 | |
| Non-lin. FTBackbone=ViT-L/14, Model formulation=f(·; θ₀ + τ)2023.05 | 85.1 | 88.8 | |
| Linear. FTBackbone=ViT-B/16, Model formulation=f_lin(·; θ₀ + τ_lin)2023.05 | 81.3 | 86 | |
| Linear. FTBackbone=ViT-B/32, Model formulation=f_lin(·; θ₀ + τ_lin)2023.05 | 76.5 | 85.4 | |
| Non-lin. FTBackbone=ViT-B/16, Model formulation=f(·; θ₀ + τ)2023.05 | 75.5 | 80 | |
| Post-hoc lin.Backbone=ViT-L/14, Model formulation=f_lin(·; θ₀ + τ)2023.05 | 75.2 | 90 | |
| Non-lin. FTBackbone=ViT-B/32, Model formulation=f(·; θ₀ + τ)2023.05 | 71.4 | 76.5 | |
| Post-hoc lin.Backbone=ViT-B/16, Model formulation=f_lin(·; θ₀ + τ)2023.05 | 65 | 85.2 | |
| Pre-trainedBackbone=ViT-L/14, Model formulation=f(·; θ₀)2023.05 | 64.4 | — | |
| Post-hoc lin.Backbone=ViT-B/32, Model formulation=f_lin(·; θ₀ + τ)2023.05 | 57.1 | 81.9 | |
| Pre-trainedBackbone=ViT-B/16, Model formulation=f(·; θ₀)2023.05 | 55.2 | — | |
| Pre-trainedBackbone=ViT-B/32, Model formulation=f(·; θ₀)2023.05 | 48.4 | — |