---
license: cc-by-4.0
task_categories:
  - other
tags:
  - brep
  - nurbs
  - cad
  - parametric-surfaces
language:
  - en
pretty_name: PARAM-BENCH
---

# PARAM-BENCH — B-Rep NURBS Surface Benchmark

**Parameterization-Agnostic B-Rep Representation Learning** benchmark data.
Built from the [ABC Dataset](https://deep-geometry.github.io/abc-dataset/) (STEP format).

> **Important**: This dataset is the *canonical, frozen* benchmark used for the
> PARAM-BENCH evaluation protocol. It is **not reproducible from ABC by
> re-running the build pipeline** (sampling is stochastic, OCC version matters).
> Use this uploaded artifact directly.

## Contents

| File | Size | Description |
|------|------|-------------|
| `surfaces.db` | 99 MB | 100,000 NURBS surfaces (SQLite, table `surfaces`) |
| `pca_augments.db` | 3.8 MB | Pre-generated augmentations (10 anchors × 50 variants) |
| `eval_protocol_canonical.json` | 58 KB | Frozen eval protocol (10 anchors / 5000 background / 300 instances) |
| `random_anchors.json` | 0.5 KB | 10 uniform-random anchors (seed=42, no type stratification) |
| `manifest.json` | 0.4 KB | Build manifest |
| `surface_vae.pt` | 29 MB | NeuroNURBS checkpoint (parameter MSE loss) |
| `surface_geom_vae.pt` | 29 MB | GeomVAE checkpoint (geometric L2 loss) |

*Note: `farthest_anchors.json` (latent-FPS anchors) is deprecated and intentionally excluded.*

## Statistics

- **100,000 faces** from 12 ABC chunks, all converted to NURBS (`BRepBuilderAPI_NurbsConvert`)
- All surfaces **clamped** (periodic surfaces unperiodized; original periodicity recorded in `is_periodic_u/v` columns)
- Dynamic stratified sampling: type × area × trim complexity
- Anchor faces: 3279, 13435, 14593, 18290, 29257, 32099, 36049, 83811, 97501, 98220

## Quick Start

```bash
# Download the whole repo (~990 MB)
pip install huggingface_hub
huggingface-cli download Axiao1999/georep-benchmark --repo-type dataset \
    --local-dir data/benchmark

# Or individual files
curl -L https://huggingface.co/datasets/Axiao1999/georep-benchmark/resolve/main/surfaces.db -o data/benchmark/surfaces.db
```

## Checkpoints

Two VAE models (identical architecture, 7.37M params, fdim=32/edim=256,
100K faces / 200 epochs / 70-20-10 split), differing **only in loss**:

- `surface_vae.pt` — **NeuroNURBS**: masked parameter MSE on control points/knots
- `surface_geom_vae.pt` — **GeomVAE**: geometric L2 on UV-sampled points (differentiable NURBS)

Expected location after download:
```
benchmark_neuronurbs/outputs/surface_vae_200ep/surface_vae.pt
benchmark_nurbs_geom/outputs/surface_geom_vae_200ep/surface_geom_vae.pt
```

## Eval Protocol

`eval_protocol_canonical.json` (PARAM-BENCH-EVAL v2.0) freezes:
- 10 uniform-random anchors (seed=42)
- 5000 background faces (real class distribution)
- 300 instances for invariance metrics
- Pre-generated augmentation database (50 variants/anchor)

## License

ABC dataset: CC-BY. Derived data distributed under CC-BY-4.0.