# Data selection and attribution

Reviewed 2026-11-02. User requested commercial-use-compatible data and exclusion
of noncommercial datasets. This recipe uses only the sources explicitly marked
included below. Repository-level labels do not replace individual asset terms.

| Source | Status | Reason |
|---|---|---|
| [scenario-labs/3d-collection](https://huggingface.co/datasets/scenario-labs/3d-collection) | Included, bounded GLB subset | Explicit CC-BY-4.0 row filter; Hunyuan-labeled models excluded; checksums verified against metadata. Attribution: Scenario; modifications: normalization, surface sampling, Gaussian proxy rendering. |
| [scenario-labs/3d-collection-hunyuan](https://huggingface.co/datasets/scenario-labs/3d-collection-hunyuan) | Excluded | Dataset expressly prohibits using assets to train or improve AI models. |
| [3dlg-hcvc/MultiScan](https://huggingface.co/datasets/3dlg-hcvc/MultiScan) | Excluded | CC-BY-NC-4.0, gated scene data. |
| [huanngzh/3D-Front](https://huggingface.co/datasets/huanngzh/3D-Front) | Excluded | CC-BY-NC-4.0, indoor scenes rather than directly aligned object targets. |
| [facebook/show3d-dataset](https://huggingface.co/datasets/facebook/show3d-dataset) | Excluded | CC-BY-NC-4.0; interaction/video data needs a separate task-specific pipeline. |
| [Voxel51/gaussian_splatting](https://huggingface.co/datasets/Voxel51/gaussian_splatting) | Deferred; not used | Card labels Apache-2.0 and describes a few scene reconstructions. This object-centered recipe does not establish camera alignment, object crops, or source-scene rights from the card alone. No importer or download is provided. |
| [allenai/objaverse](https://huggingface.co/datasets/allenai/objaverse) | Added: CC0 assets only | Dataset compilation is ODC-By; assets have separate licenses. Script accepts only individual `cc0` metadata entries, retains author/title/source links, and excludes animated assets. |
| [allenai/objaverse-xl](https://huggingface.co/datasets/allenai/objaverse-xl) | Future candidate; not used | Heterogeneous per-file licensing; no blanket commercial-use assumption or importer in this kit. |

Pinned revisions:

- Scenario: `55b4c002b4c1529fc48da3dff98516c3069bfa3f`
- Objaverse: `556637099bf4fa79ea7b239d0d9c328b8a2e9ac8`
- Original model: `1738cdfaf1bb940bf576a96d11b742ce014b9139`

The actual asset inventory is generated as `sources.json`, with asset IDs, source
URLs, authors, licenses, revision IDs and SHA-256 hashes. Published releases copy
this file to `ATTRIBUTIONS.json`. Keep it with the released model. Rejected assets
are recorded in `skipped.json`. The model card must not list excluded or merely
considered datasets as training sources.

License references: [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/),
[CC0](https://creativecommons.org/publicdomain/zero/1.0/),
[ODC-By](https://opendatacommons.org/licenses/by/1-0/).