Continual Model Merging on CLIP Task Stream 20 tasks (test)
93.5AccuracyFine-Tuned
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Fine-TunedBackbone=ViT-L/142026.05 | 93.5 | — | |
| Fine-TunedBackbone=ViT-B/162026.05 | 91.6 | — | |
| Fine-TunedBackbone=ViT-B/322026.05 | 89.8 | — | |
| ODE-MBackbone=ViT-L/142026.05 | 81.1 | -10.3 | |
| C. Fine-TunedBackbone=ViT-L/142026.05 | 77.7 | — | |
| OPCMBackbone=ViT-L/142026.05 | 76 | -6.5 | |
| Average (SWA)Backbone=ViT-L/142026.05 | 71.1 | -6.4 | |
| ODE-MBackbone=ViT-B/162026.05 | 70.4 | -9.9 | |
| OPCMBackbone=ViT-B/162026.05 | 70.3 | -6.3 | |
| C. Task ArithmeticBackbone=ViT-L/142026.05 | 70.3 | -3.3 | |
| C. Ties-MergingBackbone=ViT-L/142026.05 | 68.3 | -2.9 | |
| C. Fine-TunedBackbone=ViT-B/162026.05 | 68.2 | — | |
| ODE-MBackbone=ViT-B/322026.05 | 67.1 | -9.4 | |
| OPCMBackbone=ViT-B/322026.05 | 65.7 | -7.8 | |
| Pre-TrainedBackbone=ViT-L/142026.05 | 65.6 | — | |
| Average (SWA)Backbone=ViT-B/162026.05 | 64.8 | -7.3 | |
| C. Task ArithmeticBackbone=ViT-B/162026.05 | 64.2 | -3.6 | |
| C. Ties-MergingBackbone=ViT-B/162026.05 | 63 | -1.5 | |
| C. Fine-TunedBackbone=ViT-B/322026.05 | 62.6 | — | |
| Average (SWA)Backbone=ViT-B/322026.05 | 61.1 | -7.1 | |
| C. Task ArithmeticBackbone=ViT-B/322026.05 | 60.6 | -3.4 | |
| C. Ties-MergingBackbone=ViT-B/322026.05 | 59.9 | -1.5 | |
| Pre-TrainedBackbone=ViT-B/162026.05 | 59.8 | — | |
| Pre-TrainedBackbone=ViT-B/322026.05 | 55.6 | — |