---
pretty_name: 3DRAG-Bench
license: mit
tags:
- 3d
- glb
- mesh
- editable-3d
- 3drag
- 3d-editing
size_categories:
- n<1K
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path: metadata.csv
---

# 3DRAG-Bench

This dataset contains 93 curated 3D object assets for 3DRAG 3D editing experiments.
Each object is stored as a GLB mesh together with a cleaned editing specification.

## Dataset Structure

```text
.
+-- README.md
+-- LICENSE
+-- .gitattributes
+-- metadata.csv
+-- name_mapping.csv
`-- assets/
    `-- <asset_name>/
        +-- model.glb
        `-- dataset_input_clean.json
```

## Files

- `assets/<asset_name>/model.glb`: GLB asset file.
- `assets/<asset_name>/dataset_input_clean.json`: editing specification with handle points, target points, and confirmed editable regions.
- `metadata.csv`: tabular index for the dataset viewer and programmatic loading.
- `name_mapping.csv`: mapping from original object names to normalized asset names.

## Dataset Summary

- Number of assets: 93
- Number of GLB files: 93
- Number of JSON metadata files: 93
- Total GLB size: 138,149,824 bytes
- Total asset directory size at preparation time: 138,216,607 bytes
- Largest GLB: 8,080,936 bytes
- Smallest GLB: 598,460 bytes

## Metadata Fields

`metadata.csv` contains one row per asset:

- `asset_name`: normalized object directory name using lowercase snake_case.
- `original_asset_name`: original object name before normalization.
- `model_path`: relative path to the GLB file.
- `metadata_path`: relative path to the cleaned editing JSON file.
- `glb_bytes`: GLB file size in bytes.
- `glb_sha256`: SHA-256 checksum for integrity checks.
- `source`: source subset recorded during collection.
- `n_components`: number of editable components.
- `chosen_type`: selected editable-region representation.
- `validation_iou`: validation IoU recorded by the collection pipeline.
- `n_handle`: number of handle points.
- `n_spheres`: number of confirmed sphere regions.
- `n_cubes_center`: number of confirmed center-format cube regions.
- `n_cubes_corner`: number of confirmed corner-format cube regions.

## Naming

Asset directories are normalized to lowercase snake_case. The original object names are preserved in `original_asset_name` in `metadata.csv` and in `name_mapping.csv`.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("AeTherRaIn/3DRAG-Bench")
print(dataset["train"][0])
```

To read a GLB file after downloading the repository:

```python
from pathlib import Path

row = dataset["train"][0]
model_path = Path(row["model_path"])
metadata_path = Path(row["metadata_path"])
```

## License

This dataset is released under the MIT License. See `LICENSE` for details.

## Citation

If you use this dataset, please cite the associated 3DRAG project or paper if applicable.