--- dataset_info: - config_name: DINO features: - name: random_crop dtype: bool - name: epochs dtype: int64 - name: seed dtype: int64 - name: best_checkpoint_test_loss dtype: float64 - name: model_idx dtype: int64 - name: dataset_name dtype: string - name: best_checkpoint_test_accuracy dtype: float64 - name: weight_decay dtype: float64 - name: batch_size dtype: int64 - name: base_model dtype: string - name: best_checkpoint_val_loss dtype: float64 - name: dataset_chosen_targets dtype: string - name: best_checkpoint_train_accuracy dtype: float64 - name: best_checkpoint_train_loss dtype: float64 - name: max_train_steps dtype: int64 - name: best_checkpoint_val_accuracy dtype: float64 - name: lr_scheduler dtype: string - name: learning_rate dtype: float64 - name: random_flip dtype: bool - name: split dtype: string - name: subset dtype: string - name: hf_model_id dtype: string - name: hf_model_url dtype: string splits: - name: train num_bytes: 556679 num_examples: 701 - name: val num_bytes: 78880 num_examples: 100 - name: test num_bytes: 159344 num_examples: 201 download_size: 253029 dataset_size: 794903 - config_name: MAE features: - name: random_crop dtype: bool - name: epochs dtype: int64 - name: seed dtype: int64 - name: best_checkpoint_test_loss dtype: float64 - name: model_idx dtype: int64 - name: dataset_name dtype: string - name: best_checkpoint_test_accuracy dtype: float64 - name: weight_decay dtype: float64 - name: batch_size dtype: int64 - name: base_model dtype: string - name: best_checkpoint_val_loss dtype: float64 - name: dataset_chosen_targets dtype: string - name: best_checkpoint_train_accuracy dtype: float64 - name: best_checkpoint_train_loss dtype: float64 - name: max_train_steps dtype: int64 - name: best_checkpoint_val_accuracy dtype: float64 - name: lr_scheduler dtype: string - name: learning_rate dtype: float64 - name: random_flip dtype: bool - name: split dtype: string - name: subset dtype: string - name: hf_model_id dtype: string - name: hf_model_url dtype: string splits: - name: train num_bytes: 553950 num_examples: 701 - name: val num_bytes: 79028 num_examples: 100 - name: test num_bytes: 158815 num_examples: 201 download_size: 253409 dataset_size: 791793 - config_name: ResNet features: - name: random_crop dtype: bool - name: epochs dtype: int64 - name: seed dtype: int64 - name: best_checkpoint_test_loss dtype: float64 - name: model_idx dtype: int64 - name: dataset_name dtype: string - name: best_checkpoint_test_accuracy dtype: float64 - name: weight_decay dtype: float64 - name: batch_size dtype: int64 - name: base_model dtype: string - name: best_checkpoint_val_loss dtype: float64 - name: dataset_chosen_targets dtype: string - name: best_checkpoint_train_accuracy dtype: float64 - name: best_checkpoint_train_loss dtype: float64 - name: max_train_steps dtype: int64 - name: best_checkpoint_val_accuracy dtype: float64 - name: lr_scheduler dtype: string - name: learning_rate dtype: float64 - name: random_flip dtype: bool - name: split dtype: string - name: subset dtype: string - name: hf_model_id dtype: string - name: hf_model_url dtype: string splits: - name: train num_bytes: 559861 num_examples: 701 - name: val num_bytes: 79621 num_examples: 100 - name: test num_bytes: 160334 num_examples: 201 download_size: 254554 dataset_size: 799816 - config_name: SD_1k features: - name: model_idx dtype: int64 - name: imagenet_class_id dtype: string - name: imagenet_class_name dtype: string - name: split dtype: string - name: subset dtype: string - name: seed dtype: int64 - name: learning_rate dtype: float64 - name: max_train_steps dtype: int64 - name: rank dtype: int64 - name: pretrained_model_name_or_path dtype: string - name: n_training_samples dtype: int64 - name: hf_model_id dtype: string - name: hf_model_url dtype: string - name: hf_model_path dtype: string splits: - name: train num_bytes: 906691 num_examples: 3500 - name: val num_bytes: 64114 num_examples: 251 - name: test num_bytes: 128377 num_examples: 499 - name: val_holdout num_bytes: 67364 num_examples: 249 - name: test_holdout num_bytes: 137229 num_examples: 501 download_size: 198659 dataset_size: 1303775 - config_name: SD_200 features: - name: model_idx dtype: int64 - name: imagenet_class_id dtype: string - name: imagenet_class_name dtype: string - name: split dtype: string - name: subset dtype: string - name: seed dtype: int64 - name: learning_rate dtype: float64 - name: max_train_steps dtype: int64 - name: rank dtype: int64 - name: pretrained_model_name_or_path dtype: string - name: n_training_samples dtype: int64 - name: hf_model_id dtype: string - name: hf_model_url dtype: string - name: hf_model_path dtype: string splits: - name: train num_bytes: 924063 num_examples: 3500 - name: val num_bytes: 65187 num_examples: 251 - name: test num_bytes: 130863 num_examples: 499 - name: val_holdout num_bytes: 68302 num_examples: 249 - name: test_holdout num_bytes: 138450 num_examples: 501 download_size: 158079 dataset_size: 1326865 - config_name: SupViT features: - name: random_crop dtype: bool - name: epochs dtype: int64 - name: seed dtype: int64 - name: best_checkpoint_test_loss dtype: float64 - name: model_idx dtype: int64 - name: dataset_name dtype: string - name: best_checkpoint_test_accuracy dtype: float64 - name: weight_decay dtype: float64 - name: batch_size dtype: int64 - name: base_model dtype: string - name: best_checkpoint_val_loss dtype: float64 - name: dataset_chosen_targets dtype: string - name: best_checkpoint_train_accuracy dtype: float64 - name: best_checkpoint_train_loss dtype: float64 - name: max_train_steps dtype: int64 - name: best_checkpoint_val_accuracy dtype: float64 - name: lr_scheduler dtype: string - name: learning_rate dtype: float64 - name: random_flip dtype: bool - name: split dtype: string - name: subset dtype: string - name: hf_model_id dtype: string - name: hf_model_url dtype: string splits: - name: train num_bytes: 562795 num_examples: 698 - name: val num_bytes: 79433 num_examples: 99 - name: test num_bytes: 161793 num_examples: 201 download_size: 248900 dataset_size: 804021 configs: - config_name: DINO data_files: - split: train path: DINO/train-* - split: val path: DINO/val-* - split: test path: DINO/test-* - config_name: MAE data_files: - split: train path: MAE/train-* - split: val path: MAE/val-* - split: test path: MAE/test-* - config_name: ResNet data_files: - split: train path: ResNet/train-* - split: val path: ResNet/val-* - split: test path: ResNet/test-* - config_name: SD_1k data_files: - split: train path: SD_1k/train-* - split: val path: SD_1k/val-* - split: test path: SD_1k/test-* - split: val_holdout path: SD_1k/val_holdout-* - split: test_holdout path: SD_1k/test_holdout-* - config_name: SD_200 data_files: - split: train path: SD_200/train-* - split: val path: SD_200/val-* - split: test path: SD_200/test-* - split: val_holdout path: SD_200/val_holdout-* - split: test_holdout path: SD_200/test_holdout-* - config_name: SupViT data_files: - split: train path: SupViT/train-* - split: val path: SupViT/val-* - split: test path: SupViT/test-* tags: - probex - model-j - weight-space-learning - model-zoo - hyperparameters - stable-diffusion - vit - resnet size_categories: - 10K 🌐 Project | 📃 Paper | 💻 GitHub | 🤗 Models

![ProbeX](https://raw.githubusercontent.com/eliahuhorwitz/ProbeX/main/imgs/poster.png) ## Overview Model-J is a large-scale dataset of trained neural networks designed for research on learning from model weights. It contains **14,004** models spanning 6 subsets, each with train/val/test splits. Every row in this dataset provides the full training hyperparameters, performance metrics, and a direct link to the corresponding model weights on Hugging Face. ## Subsets ### Discriminative (one model per HF repo) | Subset | Base Model | Train | Val | Test | Total | |---|---|---|---|---|---| | **DINO** | `facebook/dino-vitb16` | 701 | 100 | 201 | 1,002 | | **MAE** | `facebook/vit-mae-base` | 701 | 100 | 201 | 1,002 | | **SupViT** | `google/vit-base-patch16-224` | 698 | 99 | 201 | 998 | | **ResNet** | `microsoft/resnet-18` | 701 | 100 | 201 | 1,002 | Each discriminative model is a full fine-tuned classifier hosted in its own Hugging Face repository. The `hf_model_id` and `hf_model_url` columns point directly to the model. ### Generative (bundled LoRA models in a single HF repo) | Subset | Train | Val | Test | Val Holdout | Test Holdout | Total | |---|---|-----|------|-------------|--------------|---| | **SD_200** | 3,500 | 251 | 499 | 249 | 501 | 5,000 | | **SD_1k** | 3,500 | 251 | 499 | 249 | 501 | 5,000 | Each generative model is a LoRA adapter. All models within a subset are bundled into a single Hugging Face repository ([SD_1k](https://huggingface.co/ProbeX/Model-J__SD_1k), [SD_200](https://huggingface.co/ProbeX/Model-J__SD_200)). The `hf_model_path` column provides the path to each model's weights within the repo. Each model's directory also contains its training images. ## Citation If you find this useful for your research, please use the following. ``` @InProceedings{Horwitz_2025_CVPR, author = {Horwitz, Eliahu and Cavia, Bar and Kahana, Jonathan and Hoshen, Yedid}, title = {Learning on Model Weights using Tree Experts}, booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)}, month = {June}, year = {2025}, pages = {20468-20478} } ```