Skip to content

Dataset sources and shared data rules

Use this page when you add or change a dataset in a model package. It lists every dataset the packages use, where the approved copy lives on Hugging Face, and the quirks that affect more than one package. Package-specific label tables, box formats, and preprocessing stay in each package README.

Dataset registry

Each model package declares the datasets it uses in its pyproject.toml under [tool.design-generators].datasets. The first column below is that declared name. When you add a dataset to a package, add its row here; tests/test_docs_generation.py fails when this table and the packages disagree. unverified in the Hugging Face column means no official dataset repository has been confirmed yet.

Metadata key Canonical dataset name Official Hugging Face dataset ID Cross-package source note
ad_banner Ad Banner unverified Current package documentation uses the original distribution until a mirror hosted by the creative-graphic-design Hugging Face organization exists.
cgl CGL creative-graphic-design/CGL-Dataset Shared poster label registry; package READMEs document package-specific use.
coco COCO unverified Current LT-Net documentation follows the original implementation's scene-graph preprocessing source.
coco-grounded COCO-Caption-Grounded unverified Current LayouSyn documentation records the original implementation's source while an import into the creative-graphic-design Hugging Face organization is unavailable.
crello Crello cyberagent/crello Approved source until a mirror hosted by the creative-graphic-design Hugging Face organization exists.
grit GRIT unverified Current LayouSyn documentation records the original implementation's source while an import into the creative-graphic-design Hugging Face organization is unavailable.
housegan-floorplan-vectorized House-GAN vectorized floorplans unverified Current package documentation identifies upstream Dropbox assets rather than an official Hub dataset.
magazine Magazine creative-graphic-design/magazine Shared layout registry; polygon inputs are converted to boxes and the source is train-only.
nsr-1k NSR-1K unverified Current LayoutGPT documentation uses original dataset JSON files.
pku_posterlayout PKU PosterLayout creative-graphic-design/PKU-PosterLayout Shared poster label registry; package-specific mappings remain in package sources.
posterlayout PosterLayout alias unverified Shared posgen alias for poster-layout data; current package sources do not establish a separate official Hub dataset ID.
publaynet PubLayNet creative-graphic-design/PubLayNet Shared layout label registry uses five document categories.
rico13 RICO13 creative-graphic-design/Rico The hosted dataset is shared with RICO25; each package preserves its original implementation's RICO13 mapping.
rico25 RICO25 creative-graphic-design/Rico Use the ui-screenshots-and-hierarchies-with-semantic-annotations configuration; the default configuration is metadata-only.
smarttext-demo SmartText demo fixture unverified Current SmartText and BASNet documentation uses local demo assets rather than an official Hub dataset.
vg-msdn VG-MSDN unverified Current LT-Net documentation follows the original Visual Genome-derived scene-graph source.
web WebUI unverified Current Parse-Then-Place documentation uses the original-implementation web pretraining split.
webui WebUI alias unverified Shared layout registry accepts this alias; no separate official Hub dataset ID is confirmed.

Shared source-selection rules

Prefer datasets hosted by the creative-graphic-design Hugging Face organization when an equivalent source exists. If no equivalent organization dataset exists, use the approved original or external source, isolate loading behind the package processor, and record a TODO with the owning model issue.

Do not fully download PubLayNet in tests; its full archive is about 107 GB. Use a builder, streaming, synthetic rows, or a tiny local fixture for ordinary checks.

Shared dataset quirks

  • RICO uses the ui-screenshots-and-hierarchies-with-semantic-annotations configuration for semantic annotations; the default configuration is metadata-only.
  • RICO13 keeps the original implementation's label mapping rather than silently reusing the RICO25 label order.
  • PubLayNet uses the shared five-category vocabulary, and package boundaries normalize source pixel boxes to the repository's public box representation where required.
  • Magazine inputs are polygon-based, converted to boxes by package processors, and train-only.
  • PKU PosterLayout includes an INVALID class in the source tensors and uses pixel ltrb boxes before package normalization.
  • CGL-v2 uses the ralf-style configuration for validation and saliency use cases.
  • Crello uses cyberagent/crello as the current public source until a mirror hosted by the creative-graphic-design Hugging Face organization exists; authors' processed splits remain the agreement baseline where a package documents them.

Import candidates

The following sources are approved candidates for future organization imports; this list does not claim that an import exists today.

  • WebUI processed data for Parse-Then-Place: issue #11
  • COCO-17, COCO-Caption-Grounded, and NSR-1K sources for Lay-Your-Scene and related work: issue #45
  • VG-MSDN scene-graph data for LT-Net: issue #12
  • Ad Banner and QB-Poster sources for LayoutDETR and PosterLLaVA: issue #18 and issue #21
  • InfoPPT processed presentation-layout data for LayoutAction: issue #13

Package-owned detail

For a package's exact dataset selection, label vocabulary, geometry, split, configuration, and evidence, read that package's Training Data section and its pyproject.toml metadata. The model package directory is the index of those package-owned sources. Use those sources for per-package tables.