Eliahu commited on
Commit
2fcce06
·
verified ·
1 Parent(s): a2f99e2

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +64 -0
README.md CHANGED
@@ -393,4 +393,68 @@ configs:
393
  path: SupViT/val-*
394
  - split: test
395
  path: SupViT/test-*
 
 
 
 
 
 
 
 
 
 
 
396
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
393
  path: SupViT/val-*
394
  - split: test
395
  path: SupViT/test-*
396
+ tags:
397
+ - probex
398
+ - model-j
399
+ - weight-space-learning
400
+ - model-zoo
401
+ - hyperparameters
402
+ - stable-diffusion
403
+ - vit
404
+ - resnet
405
+ size_categories:
406
+ - 10K<n<100K
407
  ---
408
+
409
+ # Model-J Dataset
410
+
411
+ This dataset contains the hyperparameters, metadata, and Hugging Face links for all models in the **Model-J** dataset, introduced in:
412
+
413
+ **Learning on Model Weights using Tree Experts** (CVPR 2025) by Eliahu Horwitz*, Bar Cavia*, Jonathan Kahana*, Yedid Hoshen
414
+
415
+ <p align="center">
416
+ 🌐 <a href="https://horwitz.ai/probex" target="_blank">Project</a> | 📃 <a href="https://arxiv.org/abs/2410.13569" target="_blank">Paper</a> | 💻 <a href="https://github.com/eliahuhorwitz/ProbeX" target="_blank">GitHub</a> | 🤗 <a href="https://huggingface.co/ProbeX" target="_blank">Models</a>
417
+ </p>
418
+
419
+ ![ProbeX](https://raw.githubusercontent.com/eliahuhorwitz/ProbeX/main/imgs/poster.png)
420
+
421
+ ## Overview
422
+
423
+ 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.
424
+
425
+ ## Subsets
426
+
427
+ ### Discriminative (one model per HF repo)
428
+
429
+ | Subset | Base Model | Train | Val | Test | Total |
430
+ |---|---|---|---|---|---|
431
+ | **DINO** | `facebook/dino-vitb16` | 701 | 100 | 201 | 1,002 |
432
+ | **MAE** | `facebook/vit-mae-base` | 701 | 100 | 201 | 1,002 |
433
+ | **SupViT** | `google/vit-base-patch16-224` | 698 | 99 | 201 | 998 |
434
+ | **ResNet** | `microsoft/resnet-18` | 701 | 100 | 201 | 1,002 |
435
+
436
+ 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.
437
+
438
+ ### Generative (bundled LoRA models in a single HF repo)
439
+
440
+ | Subset | Train | Val | Test | Val Holdout | Test Holdout | Total |
441
+ |---|---|-----|------|-------------|--------------|---|
442
+ | **SD_200** | 3,500 | 251 | 499 | 249 | 501 | 5,000 |
443
+ | **SD_1k** | 3,500 | 251 | 499 | 249 | 501 | 5,000 |
444
+
445
+
446
+ 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.
447
+
448
+ ## Citation
449
+ If you find this useful for your research, please use the following.
450
+
451
+ ```
452
+ @InProceedings{Horwitz_2025_CVPR,
453
+ author = {Horwitz, Eliahu and Cavia, Bar and Kahana, Jonathan and Hoshen, Yedid},
454
+ title = {Learning on Model Weights using Tree Experts},
455
+ booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
456
+ month = {June},
457
+ year = {2025},
458
+ pages = {20468-20478}
459
+ }
460
+ ```