Model-J Dataset

This dataset contains the hyperparameters, metadata, and Hugging Face links for all models in 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 | 🤗 Models

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, 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}
}