LianeMarilin/CADBench-Extended-Multimodal-Dataset

Dataset Card

Dataset Description

CADBench Extended Multimodal Dataset is an independently produced public extension for multimodal CAD reconstruction research. It contains 100 CAD samples with clean and perturbed meshes, STEP/STL/OBJ/GLB representations, single-view and four-view renders, PBR images, bilingual descriptions, prompt variants, QA, geometry metadata, grading signals, and manually reviewed visual semantics.

• Tasks: image-to-text, text-to-image, image/mesh-to-CAD research
• Languages: English and Chinese
• Public size: 100 samples; approximately 278 MiB of unpacked repository data
• Logical sample split: train 80 / validation 10 / test 10
• Source repository: GitHub

Dataset Structure and Splits

Config	Viewer split	Rows	File	Purpose
metadata	train	100	annotations/metadata_extended.jsonl	Complete nested sample metadata
rows	train	100	annotations/cadbench_rows.jsonl	Flat asset-path index
prompts	train	100	annotations/prompts.jsonl	Prompt and description variants
visual-semantics	train	100	annotations/visual_semantics.jsonl	Reviewed visual descriptions

The Viewer split labels above describe the four table entry points. The logical train/validation/test assignment is stored per sample in annotation fields and split manifests. All meshes, renders, annotations, previews, and audit files are uploaded individually, so users can browse or download selected assets without unpacking an archive.

Data Fields

metadata — sample_id, label, source, modalities, geometry, noise/render parameters, bilingual descriptions, prompts, QA, grading, and extended modalities.

rows — file_id, label, and relative paths for STL, noisy STL, OBJ, STEP, GLB, single-view, multiview, and PBR assets.

prompts — sample_id, label, logical split, bilingual descriptions, prompt variants, and visual semantics.

visual-semantics — reviewed captions, primary form, feature descriptions, feature-tree hint, confidence, and annotation method.

Loading and Asset Access

from datasets import load_dataset

rows = load_dataset(
    "LianeMarilin/CADBench-Extended-Multimodal-Dataset",
    "rows",
)["train"]
print(rows[0]["stl"], rows[0]["multiview_image"])

Use huggingface_hub.hf_hub_download(..., repo_type="dataset") with any path from the rows config to retrieve only the required asset.

Curation, Intended Uses, and Limitations

Assets are normalized into a consistent multimodal layout, rendered under recorded parameters, enriched with bilingual annotations, and checked by automated validation plus per-sample visual review. Suitable uses include CAD retrieval, multimodal representation learning, captioning, image/mesh-to-CAD evaluation, prompt studies, and pipeline QA. The public set is small, object-family coverage is uneven, synthetic renders differ from photographs, and generated perturbations do not cover every manufacturing defect. Per-asset attribution and license terms must be respected; consult annotations/license_attribution.csv and LICENSE.md.

License and Citation

The package uses per-asset licensing and attribution; Dataset Card metadata declares license: other. Preserve sample-level attribution when redistributing assets or derived data.

@dataset{cadbench_extended_multimodal_2026,
  title  = {CADBench Extended Multimodal Dataset},
  author = {Liane Marilin},
  year   = {2026},
  url    = {https://huggingface.co/datasets/LianeMarilin/CADBench-Extended-Multimodal-Dataset}
}

CADBench Extended Multimodal Dataset

CADBench Extension Data · Multimodal CAD Reconstruction · 100 Public Samples

A searchable public CADBench extension dataset with CAD meshes, four-view renders, PBR images, bilingual descriptions, prompts, splits, and QA.

Dataset Overview · Modalities · Repository Layout · Quick Start · Full Dataset

This independently produced public sample is an extension dataset for CADBench-style multimodal CAD reconstruction tasks. It follows the task structure of CADBench while providing newly prepared extension data. It is not an official CADBench release and contains no official CADBench data.

The repository is designed for searches and tasks such as CADBench extended dataset, CADBench extension data, multimodal CAD dataset, CAD reconstruction dataset, image-to-CAD, mesh-to-CAD, multi-view CAD, PBR CAD rendering, and CadQuery dataset.

---

Dataset Overview

Item	Public Release
CAD samples	100
Data files	1,800
Data size	Approximately 278 MiB
Difficulty tiers	Easy 40 / Medium 35 / Hard 25
Dataset splits	Train 80 / Validation 10 / Test 10
Provenance	Third-party open-source assets; only authors and licenses are shown
Descriptions	Chinese and English
Manual visual semantics	100 / 100 four-view composites reviewed individually
Prompts	7 prompt fields per sample
Automated validation	Passed

Modalities

Each sample includes the following standardized data:

Modality	Content	Specification
Clean Mesh	STL + OBJ	Normalized, closed manifold triangle mesh
Noisy Mesh	STL	Remeshing, vertex perturbation, and smoothing
Single-view	PNG	1280 × 1280 grayscale isometric render
Multi-view	PNG	2570 × 2570 four-view composite
PBR	PNG	1280 × 1280 randomized material and lighting
Mask / Depth / Normal	PNG	Auxiliary supervision signals
8-view	8 × PNG	Independent standardized isometric views
Text	JSONL	Geometric descriptions and reconstruction prompts

CAD Four-View Gallery

Each object is shown from four standardized isometric directions in a 2 × 2 layout. These views expose the silhouette, openings, recesses, symmetry, repeated features, and back-side geometry. The gallery below contains 12 representative Easy, Medium, and Hard samples. All 100 four-view composites are available in data/multiview/.

PBR Gallery

---

Difficulty Tiers

The extension dataset uses reproducible source triangle-count bands:

Tier	Rule	Samples	Typical Characteristics
Easy	300–1,200 triangles	40	Clear primary silhouette and relatively simple structure
Medium	1,201–4,000 triangles	35	More openings, curved regions, or repeated structures
Hard	4,001–10,000 triangles	25	Dense details, high curvature, or complex topology

These mesh-complexity tiers are part of this extension dataset and are not official CADBench difficulty labels.

---

Repository Layout

CADBench-Extended-Multimodal-Dataset/
├── data/
│   ├── clean_mesh_stl/       # 100 clean STL meshes
│   ├── clean_mesh_obj/       # 100 clean OBJ meshes
│   ├── noisy_mesh_stl/       # 100 noisy STL meshes
│   ├── singleview/           # 100 grayscale renders
│   ├── multiview/            # 100 four-view composites
│   ├── pbr/                  # 100 PBR renders
│   ├── mask/                 # 100 masks
│   ├── depth/                # 100 depth maps
│   ├── normal/               # 100 normal maps
│   ├── views8/               # 8 views × 100 samples
│   └── source_npz/           # normalized source arrays
├── annotations/
│   ├── metadata_extended.jsonl
│   ├── cadbench_rows.jsonl
│   ├── prompts.jsonl
│   ├── visual_semantics.jsonl
│   └── license_attribution.csv
├── splits/                   # train, validation, test, and difficulty lists
├── previews/                 # README galleries
├── audit/                    # schema, manifest, and validation evidence
├── tools/
│   ├── inspect_sample.py
│   ├── apply_visual_semantics.py
│   └── validate_visual_semantics.py
├── DATASET_CARD.md
└── CITATION.cff

---

Manually Reviewed Visual Semantics

Every sample was inspected individually from its full four-view composite. The review records are stored in annotations/visual_semantics.jsonl; they are not generated by a shared caption template.

Each record contains:

• a sample-specific primary-form label;
• structured openings, recesses, protrusions, repeated structures, symmetry, and surface character;
• unique English and Chinese geometric captions;
• a reviewed feature-tree hint for editable CAD reconstruction;
• annotation method and confidence.

The visual captions and structured features are synchronized into the main metadata and prompt indexes. Original mesh statistics remain available under descriptions.technical_en and descriptions.technical_zh.

Annotations and Prompts

The primary index is annotations/metadata_extended.jsonl. Each row contains:

• Sample ID, difficulty tier, split, author, and license
• Clean and noisy mesh statistics
• Relative paths for every modality
• Chinese and English geometric descriptions
• Manually reviewed visual semantics and feature-tree hints
• Rendering, camera, material, and noise parameters
• Quality-control results
• Seven task-oriented prompt fields:
  • image_to_cad_en
  • image_to_cad_zh
  • mesh_to_cad_en
  • caption_en
  • caption_zh
  • pbr_generation_en
  • negative_prompt_en

The compact prompt index is available in annotations/prompts.jsonl, while annotations/cadbench_rows.jsonl provides the CADBench extension field layout.

---

Quick Start

git clone https://github.com/gavin-sparkols/CADBench-Extended-Multimodal-Dataset.git
cd CADBench-Extended-Multimodal-Dataset
python tools/inspect_sample.py CBX100_0001
python tools/validate_visual_semantics.py --integrated

Python loading example:

import json
from pathlib import Path

root = Path("CADBench-Extended-Multimodal-Dataset")
with (root / "annotations/metadata_extended.jsonl").open(encoding="utf-8") as handle:
    samples = [json.loads(line) for line in handle]

sample = samples[0]
print(sample["sample_id"], sample["label"], sample["split"])
print(sample["descriptions"]["en"])
print(root / sample["modalities"]["multiview_image"])

Useful Entry Points

Purpose	File
Extended metadata	metadata_extended.jsonl
Prompts and descriptions	prompts.jsonl
CADBench extension fields	cadbench_rows.jsonl
Train / Validation / Test splits	splits/
Author and license attribution	license_attribution.csv
Automated validation report	validation_report.json

---

Full Dataset: 10,000+ Samples

A production-scale version with 10,000+ samples is available by request.

The full version provides broader part coverage, more categories and difficulty levels, large-scale multimodal rendering, bilingual descriptions, task prompts, quality auditing, and customizable fields.

For full-dataset access, licensing, research collaboration, or custom data production:

1. Submit the dedicated Full Dataset Access issue; or
2. Contact the repository owner @gavin-sparkols.

---

Quality and Reproducibility

The public package passed automated checks for record counts, unique IDs, difficulty quotas, dataset splits, license allowlists, file reconciliation, schema completeness, attribution records, and hidden temporary files. The manual visual-semantic layer also passes a reproducible strict check for all 100 reviewed images, unique bilingual captions, structured feature fields, feature-tree hints, and synchronization with the two public indexes.

• Validation verdict: Passed
• 100 / 100 records complete
• 1,800 / 1,800 data files reconciled
• 0 unexplained empty values
• 100 / 100 manually reviewed visual-semantic records synchronized

See audit/validation_report.json and audit/file_manifest.jsonl for machine-readable evidence.

---

Source and License

The 100 source meshes are derived from third-party open-source assets. Public metadata shows only the author, generic third-party provenance, and license type. Precise dataset names, download addresses, and original source links are omitted.

License distribution:

• Creative Commons Attribution: 87
• Public Domain: 8
• Creative Commons Public Domain Dedication: 5

Per-item author and license information is available in annotations/license_attribution.csv. Repository-authored documentation, metadata organization, and utility code follow the terms described in LICENSE.md; third-party model assets retain their original per-item licenses.

---

Citation

If this public extension dataset supports your research or project, cite this repository using CITATION.cff. Retain the author and license information recorded in annotations/license_attribution.csv when using third-party assets.

---

Search Keywords

CADBench extended dataset · CADBench extension data · CADBench dataset · CADBench GitHub · multimodal CAD dataset · CAD reconstruction dataset · image-to-CAD dataset · mesh-to-CAD dataset · CAD four-view renders · PBR CAD dataset · CadQuery reconstruction benchmark

Built for CAD reconstruction, multimodal learning, and reproducible evaluation.