Hugging Face's logo

Datasets:
JiHyuk-Byun
/
3D-PAQA

Tasks:
Other
Modalities:
Tabular
Text
3D
Formats:
csv
Size:
100K - 1M
Tags:
3d
quality-assessment
human-preference
preference-alignment
objaverse
point-cloud
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
filename
stringlengths
38
44
geometry
float64
1
5
texture
float64
1
5
material
float64
1
5
plausibility
float64
1
5
artifacts
float64
2
5
preference
float64
1
5
000-017/eee99782f3b34d1d85fabf0271c89670.npy
2.4
2.4
3
2.4
4
2.8
000-076/63abbd5997124afe8f72ce4d9014a483.npy
4.2
4
4.2
3.8
4.2
3.8
000-087/ad80f7caf008496da73a3b5e61108417.npy
3.8
4.4
3.4
4.4
3.2
2.2
000-060/7b56a35063684ca8baf8432d6d84df9f.npy
3.8
4.8
3.4
4.6
4
3.8
000-033/7976febbde784407854e8f47254f9701.npy
3.2
4.6
4
4.2
3.8
3.8
000-128/9f780649e78c4738bf0adc3a3fb8a95e.npy
2.6
4.2
3.6
1.2
4
1.6
000-121/5dbe4dc71d8841339e5e602cdcc3f1fc.npy
4.6
5
3
4.2
4.4
4.6
000-127/8cb5407a5f1a497f9bf2aa68a1ec4aac.npy
3.4
4.6
3.4
3.8
4
4
000-136/091071c1c6cb4041871554f991a8f0f8.npy
2.4
1.8
2
3
4.2
3
000-050/8a4535413b17428eab2fdc111942c269.npy
3
3.6
3.6
3.6
4
2.8
000-158/e5e9b097092e4dc0af8526f8cc37dd6d.npy
4.8
4.6
5
2.4
4
3.2
000-112/3cfb30c2c2e448dfa20b0495eeb97a45.npy
2.2
1.2
2
2
3.4
1.8
000-078/b3aae69ba5024c54adf64936ce482de5.npy
4.8
5
4.4
2
2.2
1
000-007/a219ab637f1b4889b376eb96187223f2.npy
3.6
2
2.2
3.4
3.6
2.8
000-115/dc177707fd0b43d6bd7b08a577b5af55.npy
2.6
3
3
3.2
3.8
3
000-069/72e8b8537d024e0b9bedd9197c749f43.npy
2.8
1
1.4
2.4
3.8
3
000-046/6623ded49c214b0d8ece3add7a1ad083.npy
1.2
1.8
1.8
2.4
3.4
2
000-141/8212293ab8db4baa97ce041f0d1ea60a.npy
1.6
1.4
2.4
1.6
3
1.2
000-013/2173860b4428488997dfa9e2595f40ae.npy
4.6
3.8
3.2
3
4.2
4.4
000-132/59de908347f248d599e478d0b7bdb8df.npy
3.6
2.8
2.8
2
4
2.8
000-026/bb16922146944187ab2d91268f0ab952.npy
1.6
1.8
2
3.2
4
1.8
000-066/6e49e9afd3f5423c875563aa9a188ae2.npy
2.2
2.6
1.6
2.6
3.8
2.6
000-097/76afa8c878fb403b8141d6da7547bde2.npy
2.4
3.4
2.2
3.2
3.8
3
000-017/0ed365ac5aa74c8a90563f4008919722.npy
3
2.6
1.8
2.2
3.4
2.6
000-023/af77d3fc4b3c4ade82f43ca3230241bc.npy
4.2
2.6
3
2.8
3.4
3.4
000-082/dcd7466709b44bae9127c0fd0265a16b.npy
2.6
2.2
3.8
2
4
3.2
000-029/8ccb760fa1ee46e7bde97374baf50702.npy
3.2
3.2
2.4
2.4
3.6
2.2
000-053/2e4e906bf9b447a389e5a4913bca8456.npy
2.6
3.6
1.4
1.8
3.6
1.6
000-039/c05cea3e51484331bfb4c75348d659ef.npy
3.6
5
3.6
4
4.6
4.6
000-136/ea2c231136fc4cffa32d64de02893b21.npy
3.2
4.2
1
3.2
3.6
3.4
000-007/7259f75876224248bd9104729614c213.npy
2.8
1.6
1.6
2.2
3.4
1.6
000-115/f0cbd48110ba40848a31b73211d61b66.npy
3.8
3.6
3.4
3.2
3.6
3.2
000-087/3d324c660f08404ebafe2bd4cdde3e1a.npy
4
4.8
4
2.6
3.6
4
000-107/74f6b8dd910f4aef80cd668619f0cf4b.npy
2.8
1.2
1.8
2.2
3.2
2.4
000-006/0492a8c2d1504d52a49de7e529eeab32.npy
1.6
2.4
4.6
1.8
3.2
1
000-105/40d1412aa0d84aacb6a382000ffe6523.npy
2.6
1.6
1.4
2.2
3
3
000-156/c2728d588e88463982cf3f76fe72f3f0.npy
3
4.2
3.4
3.2
4
4
000-001/6781c0c33763406ba9b95d640deef936.npy
3.2
3.6
3.8
2.8
3.8
3
000-145/49f80a0cc8e941cf9130cacf66721529.npy
3.6
4.6
2.8
3.6
3.8
4
000-020/935cc28f95f64bc4b7f0463f1f0f50e1.npy
3.4
4
2.4
2.6
4.4
3
000-059/b3bd5952ed7848358de8cdccffc6f2a8.npy
4
4.2
4.6
3
4.2
3.6
000-116/b1c12c5bcc24440690f65f2c6a9b7e83.npy
1.8
1.2
2
2.2
3.4
2.6
000-014/3dcf35af38c44e34874f5872db86cea1.npy
4
2.8
2.2
3
3.6
2.6
000-022/b3c973c1c65745e499b141c5f2cd34db.npy
5
3.6
4.6
3
4.6
3.6
000-158/baded78c9fd64d6c94ebe5082c313f25.npy
3
3.4
2.4
3
4
2.4
000-137/e5dd2babade149c7959f96b8de1618dd.npy
1.6
2.2
2
2
3.8
2.4
000-006/b1e33a99c8f84d819428c87f7b3eaee9.npy
2.8
4.8
1.4
3
4
3.6
000-097/2def35a4abaa42f3956b948121c6149c.npy
1.8
1.2
2.8
2.4
3.8
2.6
000-107/c3e5afc7eb0a46b7bb82bfbc752086d7.npy
4.6
4.2
5
4.2
4.4
4.4
000-076/df11f8090782490094aaba2bfd79819f.npy
4.8
3.6
1.8
2.6
3.4
3
000-102/9327160f834f4c7aa83084cd4e92fcb2.npy
1.4
1.6
1.4
1.6
3.8
3.4
000-087/1ea64eb458864e24932b24097d7e4c4b.npy
3
2.6
3.8
2.2
3
2.6
000-035/24bd1cd399f94575bd0084fb26f8ac9c.npy
1.8
1.6
2.2
1.2
4
1.4
000-102/93ec10e305e14f58bcd1a33e9f955fbc.npy
3
1.8
2.4
2.6
3.4
1.2
000-155/13b1ea1fb77545c5be433ce7870254ef.npy
3.8
2
3.6
2.6
3.6
2.8
000-036/f7cd9f85122c43f7b75d29d34ad1e05f.npy
3.8
4
4.2
3.4
4.2
3.4
000-021/286874f1c573415291d615c4af73118c.npy
3.2
4.2
4.2
2.4
3.8
3
000-090/b5a697b238694e01aa0bf4021f9b45bf.npy
3
4
4.6
4.2
4.2
3.6
000-034/605ed2f6247f41cc8dd70fe913612472.npy
3
2.6
1.8
2.8
3.2
2.4
000-089/d622382b490944d79fcb107938e5f4ba.npy
3.6
2.8
1.8
3.6
3.2
2
000-094/875c17326e1f47678397e4327e410c04.npy
2.6
3.4
3.8
5
5
5
000-147/4edb5fd3673041d992c08bcec314a501.npy
1.6
1
1.6
1.2
3
1.8
000-042/863dbd7b0aa54a839ccdfe43d98c778a.npy
3.4
2.8
3.8
3.2
4
4
000-131/f632eebbd9904470971ca0efd4a0c7cb.npy
4.6
5
3.8
1.4
2.8
1.6
000-144/0631239c4fbb4317a6d5b4cd77cf4622.npy
5
4.8
3
2
4
3.8
000-152/d57f243d3cbd4f5498908055e7f7ff3e.npy
3.2
2.4
2.6
2.8
3.6
2.8
000-076/3a54d7b9049f4e37b11082a4906ab81a.npy
1.6
3.6
1.8
2.2
3
2.2
000-059/267439a0666e4fa29ba86a0f03c0ffef.npy
3.2
4.8
4.4
4.2
4.2
3.6
000-134/7ddd317298714326a93a4b43c00e4ded.npy
4.2
4.8
2.4
4
4.6
4
000-137/dd1fd7e085724b3fb54d2ee6f6d2a55c.npy
4
4.4
4.4
2.8
4.2
3.2
000-095/c26df783a78f4b0cb6818d12f9a203b7.npy
4
4.4
4.6
4
3.6
3.4
000-060/d436472bfc844829afbe1d79ca58336c.npy
5
4.6
5
5
4.4
4.6
000-045/917a81110a1643a7aa3518d03ea39b3c.npy
3.4
1.2
1.2
1.6
4
2.8
000-138/07c6fa4cfa05420d89d85cf791bf8a6b.npy
2.2
1.6
1.6
2
4.2
1.6
000-143/17530a423853483b8023e95fc1a5c05a.npy
4.8
4.6
3.6
4
4.6
4
000-119/4a418d02b95449b9afae7dc39f1997fe.npy
4.4
2
1.8
4.6
3.8
2.8
000-075/0dea643703ca4887ae9d6381b1f79174.npy
1.6
5
5
1.4
4.6
3.4
000-087/6c72b3428ffd451d8b4cb8e7c9c7bfae.npy
2
2.8
3.2
1.6
2.8
1
000-079/4e368eb83b3c4f5ea8c2e1405d22d288.npy
3.2
4.8
3.8
1.4
2.8
1
000-084/45b3d4ba277b4216b1188bdcfc3fd883.npy
5
3
2.2
3.6
4.4
3
000-065/3a9b6dfc92274672888f277195e77e69.npy
2.8
5
3.4
2
2.8
1.6
000-121/0b502232b35243c8bf6a08bf49766822.npy
5
4.4
3.8
4.4
4.2
4
000-148/15f12168fbc04414bc6677ccea4ad46a.npy
5
4.2
3.4
3.6
4.2
3.4
000-050/c3492f6f06bb4472b288168d129afb6e.npy
2.2
3
1.6
2.8
3.8
2.8
000-107/9813a0a2505a464f9331ef1e2787e3b9.npy
3.6
5
4
4.2
4.2
4
000-106/c7f589d725d24e04a876dc313e2827da.npy
4.4
4
3.4
4.2
3.8
3.8
000-025/57e80a02af4e4487bf1f96eb5421f890.npy
1.6
2.6
2
1.4
3.2
2
000-153/56310456790b4b73a7abb0b5788090fb.npy
4.2
1
2.8
3.4
4.2
2.2
000-044/a514d98e3b1b48c5a54cb0cbdfdc9380.npy
2.2
4.6
1.2
2.8
3.4
4.4
000-004/6502f2d7760a416cba0c28842d0beef3.npy
3.4
3.8
4
4.4
4.4
3.8
000-147/c853345986bd4d348353db3bcd078f5b.npy
4.2
5
4.2
3.2
4
3.8
000-106/8a073df576de48d7be9a095f6b1f596e.npy
3.2
2.6
4.2
3
4.4
3.4
000-033/6835a5ba1f4c44da963942171370772c.npy
2
4.8
1.2
2.2
4
5
000-122/d3a1e710312c40289e1d09bde9520c36.npy
5
5
4.6
5
4.8
4.4
000-036/efdbdef57bc84d32b0802b8e0a8cb4f2.npy
3.8
4.8
3.8
4
3.8
3.6
000-044/2c86ffb1eef64cbd9ba71f205ec581dc.npy
1.8
3.6
1.6
2
3.6
2
000-091/9aa80bd615794b3b9d61d8adf554fa30.npy
2.6
1.6
1.4
2.2
3.8
2
000-093/6b61d246c1a447119ac6bb9190373d62.npy
5
4.6
4
4.4
4.4
4.8
000-111/a7e67c4dd64d4d4a956b5443b4d1f6ca.npy
1.4
1.2
1.8
2.8
3.8
2.8
000-139/c16369a48f6d4967b223d5bb29fa9830.npy
4.6
1.2
2.6
2.4
3.6
2.4
End of preview. Expand in Data Studio

3D-PAQA — Preference-Aligned 3D Quality Assessment

Preference-aligned perceptual quality labels for 240,636 Objaverse assets, rated on six perceptual criteria. The goal of 3D-PAQA is to move beyond synthetic-distortion 3D-QA benchmarks and provide human-preference-aligned quality scores for real, human-created 3D assets, at a scale usable for training and benchmarking automatic quality evaluators. Drawn from a 264,966-asset Objaverse corpus.

  • train.csv — 216,540 labeled assets
  • test.csv — 24,096 labeled assets

Companion model: JiHyuk-Byun/3D-PAQA-evaluator (PointTransformerV3 trained on these labels) · Training code: github.com/JiHyuk-Byun/3D-PAQA

How the labels were made

These are not raw human ratings. Collecting human MOS for hundreds of thousands of assets is not scalable, so the labels are generated by a multimodal large language model (MLLM), Qwen2-VL-72B, guided to align with human preference through a two-stage exemplar-anchored relative-ranking pipeline:

  1. Human-labeled exemplar anchors. A pool of 100 reference objects (10 categories × 10 objects) is labeled by humans and grouped into 5 quality levels (20 objects each). For each asset to annotate, one object is sampled from each level, yielding a set of 5 anchors that span the full quality spectrum.
  2. Relative Ranking (RR). The target asset (rendered to multi-view images) is presented to the MLLM alongside the 5 anchors, with criterion-specific instructions. Instead of scoring in the abstract, the MLLM ranks the target's position among the anchors per criterion; the score is derived by averaging the nearest anchors. This grounding makes outputs far more stable than direct absolute scoring.
  3. Averaging. The process is repeated 5 times with independently sampled anchor sets and averaged, reducing anchor-specific bias.

So human preference enters through (a) the human-labeled exemplar anchors injected into every prompt, and (b) a separate held-out human user study (~12k human responses on an Objaverse subset, all six criteria) that is used only to validate how well these MLLM annotations — and the evaluator trained on them — agree with people. The companion evaluator trained on these labels correlates strongly with that human study and, on the held-out subset, exceeds its teacher MLLM on every criterion.

Scope: 3D-PAQA labels cover human-created Objaverse assets. Applying it to AI-generated / text-to-3D outputs is a natural extension but is not validated by the labels here — the dataset does not (yet) contain generative-model outputs.

Schema

column description
filename <partition>/<objaverse_uid>.npy — key into the source asset
geometry mesh quality: resolution, proportions, holes/detached parts
texture base-color / albedo quality
material metallic-roughness / PBR quality
plausibility overall realism / how convincingly it depicts the object
artifacts freedom from self-intersections, degenerate faces, flipped normals, stray geometry (higher = fewer artifacts)
preference overall preference across all criteria

All scores are on roughly a 1–5 scale (higher = better) and are asset-level (they apply to the mesh; the .npy suffix refers to the processed point-cloud form below).

from datasets import load_dataset
ds = load_dataset("JiHyuk-Byun/3D-PAQA")
print(ds["train"][0])
# {'filename': '000-015/3876fb82991240c7bf4856c4bb4185b0.npy',
#  'geometry': 4.0, 'texture': 4.4, 'material': 4.8,
#  'plausibility': 3.2, 'artifacts': 3.6, 'preference': 2.8}

Normalization — labels are raw on purpose

The labels are shipped un-normalized (raw scores). This is intentional: the raw score is the supervision target, and the companion evaluator regresses these raw values. The six criteria are on different effective scales (train split):

criterion mean std p05 p50 p95
geometry 3.22 1.04 1.4 3.2 5.0
texture 3.37 1.27 1.0 3.8 5.0
material 2.82 1.12 1.2 2.6 4.8
plausibility 2.92 0.89 1.6 2.8 4.6
artifacts 3.79 0.49 3.0 3.8 4.6
preference 2.96 0.88 1.4 3.0 4.4

If you want to normalize downstream (e.g. to compare criteria on one scale, or to balance a multi-task loss), use criteria_stats.json (per-criterion mean/std and a cdf_grid computed on the full train split):

import json
s = json.load(open("criteria_stats.json"))["stats"]
z = (x - s["material"]["mean"]) / s["material"]["std"]   # z-score

Note z-scoring is affine — it rescales but does not change each criterion's distribution shape (some criteria, e.g. artifacts, are skewed/peaked).

Getting the 3D assets (Objaverse provenance)

This repo distributes labels + processing code only. The underlying 3D meshes are from Objaverse and remain under their original per-asset licenses, so we do not redistribute them. Reproduce the assets yourself from the UID in filename:

pip install objaverse trimesh numpy
python process_objaverse.py --csv train.csv --out ./pc_npy

process_objaverse.py downloads each asset by its Objaverse UID and runs mesh2pc.py to produce the exact 11-feature point cloud the evaluator consumes: per point [coord(3), color(3), normal(3), metallic(1), roughness(1)]. This sampler is the data contract used at label/train time — keep it byte-identical.

License & attribution

  • Labels & code in this repo: CC-BY-4.0 (this work).
  • Source 3D assets: Objaverse, each under its own license — not included here; download via the script above and respect per-asset terms.

Citation 

@misc{byun2026_3dpaqa,
  title  = {3D-PAQA: Towards Preference-Aligned 3D Quality Assessment},
  author = {Byun, JiHyuk},
  year   = {2026},
  note   = {https://huggingface.co/datasets/JiHyuk-Byun/3D-PAQA}
}
Downloads last month
111
Number of rows:
240,636
Total file size:
16.6 MB

Models trained or fine-tuned on JiHyuk-Byun/3D-PAQA