README.md

---
pretty_name: 3DVerifierBench
language:
  - en
license: other
license_name: 3dverifierbench-mixed-scope
license_link: https://huggingface.co/datasets/AI-YesLab/3DVerifierBench/blob/main/licenses/README.md
task_categories:
  - visual-question-answering
size_categories:
  - 1K<n<10K
tags:
  - 3d
  - verification
  - benchmark
configs:
  - config_name: benchmark
    default: true
    data_files:
      - split: test
        path: metadata/benchmark.jsonl
---

# 3DVerifierBench

**Grounding Generative Verification in 3D Content**  
Shuanglin Li, Jiaying Lin, Shuquan Ye.

[Paper: arXiv:2610.22360](https://arxiv.org/abs/2610.22360)

The test set contains **2,581 samples across ten tasks**, with 970 True and
1,611 False answers. It includes the original image/video inputs and available
3D assets. Related samples are grouped into 1,167 cases.

## Test annotations

`metadata/benchmark.jsonl` contains one sample per line, with eight fields:

| Field | Meaning |
| --- | --- |
| `sample_id` | Short neutral ID, such as `object_0001_01` |
| `task_id` | Task name: `object`, `part`, `material`, `part_grounding`, `part_relation`, `affordance_region`, `grasp`, `identity`, `fine_geometry` or `motion` |
| `question` | The factual claim; for Motion, the complete caption-based instruction |
| `answer` | Ground-truth Boolean judgment |
| `reference_explanation` | Reference rejection explanation used for scoring |
| `instruction` | Static-task instruction; Motion's instruction is already in `question` |
| `input_type` | Original image/video input protocol |
| `evidence` | Ordered image/video entries with paths, media IDs and content hashes |

IDs follow `<task>_<case>_<sample>`, matching the evidence directory. For example,
`grasp_0090_01` uses `evidence/grasp/case_0090/sample_01/`. Sample numbers use a
fixed neutral ordering and do not encode the answer. Each sample is evaluated
independently. All paths are relative to the dataset root.

`metadata/inputs.jsonl` is the same public input projection without answers or
reference explanations. Object, Part and Material tasks use the question and eight ordered views;
their generic task instruction is omitted from model input, following the paper.
`metadata/motion_frames16_inputs.jsonl` supplies the 134 alternative sixteen-frame
Motion inputs with the original sampling and layout information.

Evaluation grouping, sample construction type, source attribution and geometry
availability are in `metadata/sample_metadata.jsonl`. The exact source/candidate
3D asset associations are in `metadata/sample_asset_links.jsonl`. Join these
indexes by the same short `sample_id`; asset filenames have independent numbering.
The code release's reader joins the indexes automatically. Original questions,
answers, reference explanations, media bytes and evidence order are preserved.

## Directory layout

```text
3DVerifierBench/
├── README.md
├── metadata/
│   ├── benchmark.jsonl
│   ├── inputs.jsonl
│   ├── motion_frames16_inputs.jsonl
│   ├── sample_metadata.jsonl
│   ├── sample_asset_links.jsonl
│   ├── grasp_poses.jsonl
│   └── dataset_stats.json
├── evidence/
│   └── <task>/case_0001/
│       ├── source.glb
│       ├── candidate_01.glb
│       └── sample_01/
│           ├── view_01.png
│           └── ...
├── manifests/
│   ├── benchmark_manifest.json
│   ├── pair_links.jsonl
│   └── triplet_links.jsonl
└── licenses/
```

The tasks are Object, Part, Material, Part Grounding, Part Relation, Affordance
Region, Grasp Validity, Identity Preservation, Fine-grained Geometry and Motion.

## Images, videos and 3D assets

- Object, Part and Material tasks have eight ordered images, `view_01.png` through `view_08.png`.
- Region, relation, affordance and Grasp use `views.png`, `annotated_global.png`
  and `annotated_detail.png`.
- Identity and FineGeo use `pair.png`.
- Motion has `video.mp4`, the alternative `frames16.png`, and original OBJ frames
  under `mesh_sequence/frame_00001.obj`, with `captions.txt`.
- Identity assets include A/B OBJ and PLY meshes, source views and synthesis data.
- FineGeo includes supplied candidate GLBs and official `reference_cad.ply` files.
  `metadata/fine_geometry_models_info.json` records their upstream CAD metadata.

There are 11,769 image/video input files, representing 7,681 unique original
contents, plus 12,420 source asset files and 200 complete Grasp scene GLBs. The
134 Motion sequences contain 6,088 original OBJ frames.

Every Grasp sample contains `scene.glb` with its unchanged object, palm and both
fingers at the recorded candidate pose. Its case-level `source.glb` remains the
object alone. `metadata/grasp_poses.jsonl` retains the 4x4 source transform and
gripper dimensions in meters. The complete scene applies the same canonicalization
to object and gripper, then converts source Z-up to glTF Y-up. Rebuild the scenes
from the code checkout with:

```bash
python scripts/export_grasp_scenes.py --benchmark-root /path/to/3DVerifierBench
```

Forty DB-3DME samples have original turntable GIFs instead of source meshes.
Thirty FineGeo samples have official reference CADs without an original candidate
mesh. Their `geometry_status` is recorded in `metadata/sample_metadata.jsonl`.
Provided images/videos remain authoritative for reproducing the paper's visual
input protocol; new renders may have different cameras and pixels.

## Load

```python
from datasets import load_dataset

test = load_dataset("AI-YesLab/3DVerifierBench", split="test")
print(test[0]["sample_id"], test[0]["question"], test[0]["answer"])
```

Download the complete repository to resolve evidence and 3D asset paths locally:

```python
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(repo_id="AI-YesLab/3DVerifierBench", repo_type="dataset"))
sample = test[0]
views = [root / item["path"] for item in sample["evidence"]]
```

For an already extracted ZIP, use the local JSON reader:

```python
test = load_dataset("json", data_files={"test": str(root / "metadata/benchmark.jsonl")}, split="test")
```

Install the reader with `python -m pip install datasets`. Source meshes, labels,
reference explanations and related-sample metadata are excluded from subject
prompts. The separate `3DVerifier` code release handles prompt construction,
model execution and scoring.

## Evaluate

From the code checkout, configure your own model endpoint and credentials:

```bash
python scripts/prepare_manifest.py --benchmark-root /path/to/3DVerifierBench --task part --output work/part_manifest
python scripts/run_evaluation.py --manifest work/part_manifest --config examples/direct.json --output runs/part
python scripts/score_run.py --run runs/part
```

For native Motion, prepare its video manifest with
`scripts/prepare_motion_manifest.py` and use `scripts/evaluate_motion_video.py`.
Preserve the model's original native-video or sixteen-frame protocol for paper
reproduction.

## Indexes, license and citation

`manifests/benchmark_manifest.json` records paths, membership and content hashes.
`metadata/dataset_stats.json` provides global and per-task counts. The linked
indexes contain 661 specification pairs, 727 candidate pairs and 615 triplets.
The benchmark's 1,123 recorded source groups are disjoint from the training
resource's 4,011 groups; this check does not cover unrecorded semantic duplicates.

Original annotations are CC BY 4.0. Upstream media, meshes, captions and native
annotations retain their source-specific terms. See
[licenses and attribution](licenses/README.md) and retain the supplied notices.
Training data and model weights are separate distributions.

Please cite the [arXiv preprint](https://arxiv.org/abs/2610.22360):

```bibtex
@misc{li2026groundinggenerativeverification3d,
  title={Grounding Generative Verification in 3D Content},
  author={Shuanglin Li and Jiaying Lin and Shuquan Ye},
  year={2026},
  eprint={2610.22360},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.22360}
}
```