KOLCO CLUBDesign training data

Datasheet

Datasheet for the KOLCO design corpus

Structured after Datasheets for Datasets (Gebru et al., 2018): motivation, composition, collection, preprocessing, uses, distribution and maintenance. Prepared for model-card and compliance review, and revised with each release tag.

Motivation

Purpose and origin of the corpus

The corpus was produced as a commercial library of editable design templates and is licensed for model training as a second use. Because every file was made to be opened and edited by a person, the text is live, the structure is intact and nothing is flattened.

Producer. The Doodle Desk, named as dc:creator in the embedded metadata of the illustration line and as producer in every catalogue index.

Publisher and licensor. KOLCO COMMERCE PRIVATE LIMITED, operating KOLCO CLUB. The contracting party on every licence.

Composition

Collections, class axes and labels

Counts are shipping counts from the catalogue indexes; sample measurements are labelled as such.

AxisPoster & printInfographicsFloor plans
Primary classes205 business verticals56 sectors5 unit types
Secondary classes587 layout families896 subjects6 footprints, 6 circulation types
Colour519 palettesPalette per file16 palettes
Type56 typographic themesTypefaces per file8 typefaces
Format23 print formats16 page formats3 dimension systems
Other7 product types417 towns and cities25 symbol styles

Floor-plan distributions

UnitPlans
3 BHK153,454
2 BHK72,372
4 BHK54,955
1 BHK15,315
Studio3,904
FootprintPlans
L-shaped209,879
T-shaped65,694
Rectangular17,600
Stepped4,015
U-shaped1,485
Notched1,327

Metadata schema

Poster & print: line, folder, file, width_px, height_px, text_nodes, path_elements, kb, with embedded Dublin Core (dc:title, dc:creator, dc:subject, dc:type, dc:coverage). Infographics: line, file, page_size, width, height, subject, sector, palette, typefaces, dark, size_kb. Floor plans: 21 fields; a real row is shown on the floor plans page.

Labels as shipped

FileSourceLabels
60036_spa_festive_flyer_a4.svgEmbedded RDFdc:title Day Spa Poster — Season’s pick: massages · dc:type Poster · dc:coverage Character Scene Poster · dc:subject beauty, day-spa, editable-text, festive, flat, flyer_a4, poster, vector
OV100109_poster_big_type.svgEmbedded RDFdc:title Halloween Event A3 Poster — Bold Type with Cartoon Gnu Art · dc:coverage A3 Portrait Poster · dc:subject a3-p, bold-type, cartoon, dark, editable-text, gnu, halloween-event, poster, vector
1254792_app_Brightmere-Group_2018.svgCSV indexline Slide-deck · subject app · Brightmere Group · 2018 · page_size 16:9 Widescreen, 1962 × 1104 · typefaces Figtree
1252005_fleet_Fold-Street_2013.svgCSV indexline Slide-deck · subject fleet · Fold Street · 2013 · typefaces Gudea
54005626_furnished.svgCSV indexunit STUDIO · 369 sq ft · 5 rooms · 1 toilet · 1 balcony · footprint L · circulation open distribution · facing N
300019524_furnished.svgCSV indexunit STUDIO · 480 sq ft · 6 rooms · footprint L · circulation central lobby · facing W

Collection

How the files were produced

Every asset was composed by people from vector artwork and typographic systems that people built, inside one production pipeline per collection. There is no crawl, no third-party aggregation and no machine authorship anywhere in the corpus.

Single pipeline per collection. Each collection is one production line with one schema, which is why the metadata is uniform across hundreds of thousands of files.

Human authorship. No machine authored any part of any file. The licence note that accompanies each pack states this.

Illustrative content. No file depicts a real person, organisation, building or measurement. Figures on a page are written by the producer and are internally consistent.

Self-contained files. No file references a font, image or asset it does not carry. Typefaces are embedded under a libre licence or named as system faces with fallbacks. Verified across the public sample: zero remote references in 191 files.

Preprocessing

Checks performed on the public sample

The checks below were run by parsing the 191-file public sample directly. Parse validity, live text, self-containment and size are measured, not asserted, and the sample allows the measurements to be repeated.

CheckPostersInfographicsFloor plans
Files in sample924851
Parse as valid XML92 / 9248 / 4851 / 51
Carry live, non-empty text92 / 9248 / 4851 / 51
Reference a remote font or asset000
Median file size88 KB76 KB41 KB
Median live text nodes761151
Median vector paths2032275

Uses

Recommended and not recommended

Recommended. Image-to-SVG, text rendering with ground truth, layout and hierarchy, chart reasoning, floor-plan understanding and conditional layout. Any task where an aligned vector source is the label.

Not recommended. Any task that treats the figures, names or locations on a page as facts about the world. They are illustrative. The corpus contains no photographs and no depictions of real people, and should not be used to infer either.

Distribution

Formats, delivery and terms

ItemDetail
FormatsSVG sources; PNG renders at 512, 1024 and 2048 px; per-file CSV metadata; embedded Dublin Core RDF on the illustration line
PackagingWebDataset shards or Parquet; manifest with SHA-256 per file
DeliveryS3 or GCS bucket, private Hugging Face repository, or signed archive links
LicenceAnnual, non-exclusive training licence per release tag; agreement supplied before any commercial discussion

Maintenance

Versioning and withdrawal

Every release carries an immutable tag. Additions ship quarterly as new tags. A withdrawn file is listed explicitly in the next tag’s manifest, so a licensee can reconcile a training set without comparing archives. This datasheet is revised with each tag and the revision is recorded in the manifest.