Presentation Slide Dataset

Overview

What this dataset is

• Live text
• Chart families in slot order
• Deck template family
• Grid and palette IDs
• Content digest per page

AI use cases

• Slide generation
• Layout generation
• Chart understanding
• Text-to-SVG
• Design AI
• Evaluation

Specifications

Dataset specification

Format, DOM node types, interleaved text and graphics, and the content rules (no brands or trademarks, no personally identifiable information) are stated for every dataset on the specifications page.

Production lines

Taxonomy coverage

Deck template families

306

Chart families

131

Grids

33

Palettes

306

Illustrations

2,155

Typefaces

120 families

Schema

Per-asset record

Parquet, mirroring the field conventions of MMSVG-2M and Hugging Face datasets so existing loaders work unchanged. The texts[] and layout[] arrays are not offered by any public SVG dataset.

asset_id, dataset_id, dataset_version, line, file_path, sha256
svg_raw
svg_normalized       fixed viewBox, transforms baked, CSS inlined, M/L/C/Q/A/Z only
png_448, width, height, orientation, page_format
text_count, path_count, group_count, image_count, element_count, token_len, complexity_tier
texts[]              {content, role, font_family, font_weight, font_size, bbox, script}
layout[]             {element_id, type, bbox, z_order, parent_group}
palette_id, colours[], is_dark, font_ids[], font_licences[]
sector, subject, industry, geography, tags[], headline
caption_short, caption_medium, caption_detailed
provenance           {template_id, illustration_source_url, illustration_licence, generator, generated_at}
quality              {valid_svg, renders, has_live_text, has_raster, near_dup_group, phash}

Quality

Checks and results

Badges are published now. A composite SVGZO Quality Score follows once the formula is frozen and applied to every line. Method →

Provenance

Where the data comes from

• Assets are composed, created and processed into SVG from source files by The Doodle Desk's team and network of creators.
• Every figure, name and caption is a written placeholder. No real organisation, person or measurement appears.

A copy-ready EU AI Act training-summary paragraph and a per-asset manifest ship with every licence. Trust and provenance →

Licence

Tiers available for this dataset

Full texts on the licensing page. Drafts pending counsel review.

Files

Loading the data

Machine-readable metadata: croissant.json. Version history: changelog.

from datasets import load_dataset

ds = load_dataset("parquet", data_files="metadata/*.parquet", split="train")
row = ds[0]
print(row["text_count"], row["path_count"], row["caption_short"])

# Full SVG files ship as tar shards; verify before extracting:
#   sha256sum -c checksums/SHA256SUMS

FAQ

Common questions

Is the content real?

No. Every name, figure and caption is a written placeholder. The dataset teaches layout, chart construction and typography, not real-world statistics.

Where does the data come from?

Assets are composed, created and processed into SVG from source files by The Doodle Desk's team and network of creators. A per-asset manifest and a training-content summary paragraph ship with every licence.

Can I train a commercial model?

Yes, under the Commercial AI or Enterprise tier. The Research tier is limited to non-commercial deployment.

Can I redistribute the files?

No tier permits redistributing or reselling the raw data. Models trained on it are yours.

How is it delivered?

Sharded tar archives with SHA-256 sums, a parquet metadata index and a Croissant manifest, via signed object-storage URLs or a scoped bucket for rclone.