---
library_name: pytorch
tags:
- cad
- jepa
- graph-representation-learning
---
# JEPA-CAD pretrained comparison artifacts

Exact existing Windows artifacts for the frozen `jepa_joint` comparison arm.
No preprocessing, retraining, or new result generation was performed for this upload.

| Artifact | Bytes |
| --- | ---: |
| `histcad_large_joint_jepa_v8/data/processed/full_paired_graphs_v7.pt` | 8,075,724,068 |
| `histcad_large_joint_jepa_v8/outputs/full_dataset_4096/best.pt` | 1,938,955,953 |

The cache contains 162,131 prepared HistCAD graph pairs with the original splits
(145,915 training / 8,130 validation / 8,086 test). The pretrained checkpoint was
selected at epoch 20 and produces a 4,096-dimensional joint representation.

Code and results: https://github.com/OliverFanHL/JEPA-CAD

Original HistCAD source dataset: https://huggingface.co/datasets/DongXintong/HistCAD
The graph cache is a processed derivative; original STEP files, parametric JSON
source files, and source dataset archives are not included.

## Download on Linux

From the JEPA-CAD code checkout:

```bash
python -m pip install --upgrade huggingface_hub
hf download IELTS8/JEPA-CAD histcad_large_joint_jepa_v8/data/processed/full_paired_graphs_v7.pt histcad_large_joint_jepa_v8/outputs/full_dataset_4096/best.pt --local-dir .
python use_pretrained.py
```

`ARTIFACTS.json` records SHA-256 checksums. The checkpoint and graph cache are
PyTorch serialized artifacts for use with the matching code. No optimizer step
or preprocessing is needed to use this frozen comparison arm. The code repository
also provides a checksum-verifying downloader with a pinned revision.