Model-J: MAE Model (model_idx_0270)

This model is part of the Model-J dataset, introduced in:

Learning on Model Weights using Tree Experts (CVPR 2025) by Eliahu Horwitz*, Bar Cavia*, Jonathan Kahana*, Yedid Hoshen

🌐 Project | 📃 Paper | 💻 GitHub | 🤗 Dataset

Model Details

Attribute	Value
Subset	MAE
Split	val
Base Model	`facebook/vit-mae-base`
Dataset	CIFAR100 (50 classes)

Training Hyperparameters

Parameter	Value
Learning Rate	0.0003
LR Scheduler	cosine_with_restarts
Epochs	2
Max Train Steps	666
Batch Size	64
Weight Decay	0.007
Seed	270
Random Crop	False
Random Flip	False

Performance

Metric	Value
Train Accuracy	0.8579
Val Accuracy	0.7653
Test Accuracy	0.7608

Training Categories

The model was fine-tuned on the following 50 CIFAR100 classes:

orange, cup, sweet_pepper, lizard, rocket, can, apple, bicycle, willow_tree, beetle, crab, plate, television, forest, snail, streetcar, bed, wardrobe, sunflower, caterpillar, seal, lamp, chimpanzee, mouse, mushroom, clock, tank, shark, raccoon, bowl, bear, wolf, tulip, crocodile, maple_tree, man, plain, butterfly, porcupine, whale, bottle, trout, dolphin, telephone, rabbit, bridge, mountain, possum, otter, oak_tree