Draw2Explain

Draw2Explain is an object-centric engineering illustration dataset for generating cutaways, cross-sections, and other structure-explaining technical drawings. The released training split contains 150,720 image-caption pairs in 151 WebDataset shards. Each image keeps its source URL, creator, attribution, per-item rights metadata, and SHA-256 digest.

Representative examples

These are actual records from the released training split, not benchmark cases. The text below each image is the sample's caption_short_en; complete bilingual captions, generation prompts, component constraints, and relation constraints are stored in the paired .json member.

Pratt & Whitney R-4360 engine cutaway · Level 4 commons:107825147 · source · CC0	A detailed cutaway of a turbofan engine highlighting key components. · Level 4 commons:12365146 · Richard Wheeler (Zephyris) · source · CC BY-SA 3.0
Exploded view of a digital camera showing internal components. · Level 4 commons:12030060 · Shigeru23 · source · CC BY-SA 3.0	Cutaway illustration of the Lunar Module. · Level 4 commons:14623428 · NASA-MSFC / Adert · source · public domain
Cross section of the steamship “Baltic”. · Level 5 internet_archive:everettsencyclop00ever:n289 · source · public domain	Sectional view of a turbine showing the electric generator above. · Level 5 internet_archive:pickshovelpluckf00bond:n171 · source · public domain
Cross-sectional view of a four-cylinder internal combustion engine. · Level 5 internet_archive:ourwonderworldal02unse:n145 · source · public domain	Evans-Klepetko furnace details. · Level 5 internet_archive:cu31924004606608:n117 · source · public domain

The target visual semantics are: a recognizable physical engineered object, normally hidden internal structure, meaningful component relationships, and an explanatory technical presentation. Ordinary product photos, software diagrams, flowcharts, plots, circuit schematics, and decorative technical-looking art are outside the intended distribution.

中文简介

Draw2Explain 面向结构解释型工程图生成，重点覆盖剖视图、剖面图和展示内部 构造的技术插图。发布版包含 150,720 组图像—标注对；每条标注 提供中英文描述、英文生成提示词、可见部件与空间/机械关系约束，以及来源、 许可和哈希信息。标注由视觉语言模型生成并通过结构与绑定校验，目前仍属于 “待独立人工图像核验”状态；使用者必须按每条样本自己的许可字段履行署名或 相同方式共享等义务。

What is included

Each sample has two members with one shared key:

• .image: the original encoded image bytes. The original extension and MIME type are recorded in metadata.
• .json: English/Chinese descriptions, a generation prompt, component and relationship contracts, image-grounded evidence, layout attributes, source provenance, license metadata, and integrity hashes.

Important fields include prompt_t2i_en, caption_detailed_en, caption_detailed_zh, required_elements, required_relations, component_evidence, relation_evidence, visible_text, layout_attributes, source_page_url, license_tier, license_url, image_sha256, and caption_prompt_sha256.

Loading

from datasets import load_dataset

ds = load_dataset(
    "webdataset",
    data_files={
        "train": "hf://datasets/Tuyuanpeng/Draw2Explain/data/train-*.tar"
    },
    split="train",
    streaming=True,
)

sample = next(iter(ds))
image_bytes = sample["image"]
metadata = sample["json"]
prompt = metadata["prompt_t2i_en"]

The neutral .image suffix deliberately keeps the WebDataset column schema uniform across JPEG, PNG, WebP, GIF, TIFF, and BMP sources. Decode the bytes with Pillow or your training pipeline using original_image_extension/mime_type.

Caption construction and status

Captions were produced with Qwen2.5-VL-7B-Instruct at revision cc594898137f460bfe9f0759e9844b3ce807cfb5 using deterministic decoding from the image plus hash-bound source hints. Outputs were normalized only for schema-preserving format repairs, then checked for exact ID/hash binding, required field structure, length constraints, component/relation evidence coverage, and duplicate text. Quality rejections were excluded from this release.

These captions are machine-generated and pending independent human audit. They must not be described as human-verified annotations. The release exposes this status in caption_review:

Value	Samples	Share
vlm_image_grounded_pending_human_audit	150,720	100.00%

Data distribution

Providers

Value	Samples	Share
uspto_patents	58,508	38.82%
google_patents_us	52,984	35.15%
internet_archive	34,586	22.95%
wikimedia_commons	4,180	2.77%
openverse	243	0.16%
library_of_congress	142	0.09%
smithsonian_open_access	56	0.04%
wellcome_collection	13	0.01%
nasa_images	8	0.01%

License tiers

Value	Samples	Share
public_domain	147,859	98.10%
sharealike	1,822	1.21%
attribution	541	0.36%
cc0	498	0.33%

Visual match levels

Level 5 denotes a rich engineering explanation plate, level 4 a strong object-centric cutaway/cross-section, and level 3 a useful but simpler structural illustration. model_audited identifies accepted records awaiting assignment to one of the numbered human-review levels.

Value	Samples	Share
3	97,459	64.66%
unknown	35,852	23.79%
4	17,338	11.50%
5	71	0.05%

Rights and attribution

This is a mixed-license dataset, so the repository-level license is other. Rights must be evaluated per sample using license_tier, license_raw, license_url, creator, and attribution. The collection includes public-domain, CC0, attribution-required, and share-alike material. Users are responsible for following the applicable attribution and share-alike terms; this card is not legal advice.

Integrity and benchmark isolation

• Every released image passed decode checks during curation and was SHA-256 rechecked while the WebDataset shards were built.
• IDs and image hashes are unique in the frozen training manifest.
• A separate 300-case benchmark is withheld from training.
• Training/benchmark leakage checks cover record ID, exact image hash, source family, and confirmed perceptual duplicates.
• SHARDS.json records every shard's byte size, sample count, and SHA-256 digest.

The benchmark is not included in this training release. Its reference images and contracts remain gated on two independent review files before public benchmark publication; this prevents unreviewed evaluation targets from being presented as final ground truth.

Limitations

The distribution is dominated by public-domain patent drawings and therefore is not a balanced census of all engineering illustration styles. Fine printed text, small callout numbers, and visually ambiguous parts can be transcribed or named incorrectly. Source metadata may be incomplete, and machine-generated captions can contain grounding errors despite schema and evidence checks. Users should run a task-specific human audit before safety-critical, educational, or commercial deployment.

Citation

Paper and citation metadata will be added with the benchmark paper. Until then, please cite this Hugging Face dataset repository and the exact revision used.