KOLCO CLUB · Training data for design models
Three million editable SVG design files are shipping today, each with live text and its own metadata row. Creators add to the catalogue and earn every time it is licensed. AI companies license a slice of it — with consent and provenance attached to every single file.
KOLCO CLUB has always been a creator commerce platform. This is the same idea pointed at AI: the people who draw the work get paid for it, and the people training models get data they can actually account for.
For creators
The folders of vectors, icon sets and illustrations you have already drawn can earn on every release they appear in — without giving up the right to sell them anywhere else.
For AI companies
Vector-native design data with a documented consent chain, so your provenance section is a table of facts rather than a paragraph of hedging.
These are shipping collections from The Doodle Desk, not a roadmap. Every file is real SVG with live editable text — no outlined type, no flattened artwork — and every asset carries its own metadata row, so a lab can license a slice instead of running a scrape.
Single-page print products at known trim sizes — the headline, body copy, call to action and contact line all live text.
Charts, rankings and breakdowns where the arithmetic on a page agrees with itself — before and after you change a figure.
Indian residential plans, studio to four bedroom, with every room name, dimension, area figure and schedule row as live text.
Volume is easy to buy and worthless on its own — a million near-identical pages teaches a model one page. What a design model needs is spread: across business context, layout, palette, format and drawing convention. These are the axes the catalogue is actually built on, and the counts are the shipping counts.
These distributions are uneven on purpose, and we would rather you saw it here than found it after signing. Three-bedroom plans and L-shaped footprints dominate because that is what gets built; carpet areas cluster hard between 1,000 and 1,249 sq ft, with a median of 1,123. If your model needs a flat distribution instead of a realistic one, say so — a release can be sampled to a target balance, or a commissioned set can fill the thin classes.
Every claim below was run against the 191 SVG files in the public review sample — stratified, drawn at random within each format, and not retouched. Download the sample and reproduce the whole table.
| Check | Posters | Infographics | Floor plans |
|---|---|---|---|
| Files in sample | 92 | 48 | 51 |
| Parse as valid XML | 92 / 92 | 48 / 48 | 51 / 51 |
| Carry live, non-empty <text> | 92 / 92 | 48 / 48 | 51 / 51 |
| Reference a remote font or asset | 0 | 0 | 0 |
| Median file size | 88 KB | 76 KB | 41 KB |
| Median live text nodes | 7 | 61 | 151 |
| Median vector paths | 20 | 32 | 275 |
Fonts travel inside the file as data URIs, or are named system faces with generic fallbacks. Nothing reaches outside itself for a font or an image, so a file renders identically wherever it lands — including inside a training pipeline that has no network.
Nothing below was written for this page. Two of the posters carry Dublin Core RDF inside the SVG itself — title, creator, type, coverage and subject tags — and the rest are read straight from the CSV index that travels with each pack. This is what a training pair looks like before you touch it.
Labels come from two places and both travel with the file. The illustration line writes Dublin Core RDF into every SVG, so the caption survives even if the file is moved out of the pack; every collection also ships a CSV index with one row per asset on the schema shown above. Where a caption needs to be richer than the tags — a written description for image–text training, say — that is a commissioned pass we can run over any release.
Every asset ships with its own metadata row. The artwork is one field of it; the rest is what makes the artwork searchable, filterable and trainable. This is a real row from the floor-plan index, not a schema sketch.
The difference between a crawl and a catalogue is that someone chose to be in the catalogue. Every file here arrived because a creator uploaded it and agreed to the terms attached to it.
No generative model produced any part of any file in these packs. Every page was composed from vector artwork and typographic systems that people built. For anyone training a design model, that means the data is not quietly recycling another model’s output.
Creators grant training rights per submission under a named, versioned agreement. No blanket clause buried in a sign-up flow. If the grant changes, they re-sign it or their files stay out of the next release.
No file references a font, an image or an asset it does not carry. Verified across the whole review sample: zero remote references in 191 files. Faces are either embedded as data URIs under a libre licence, or named system faces with generic fallbacks, so no font licence travels with a file.
No page describes a real organisation, a real building or a real measurement. Names, figures and areas are written placeholders that are internally consistent — a page agrees with itself — and none of it should be presented, or trained on, as fact.
Four steps from a folder on your desktop to a line in your payout statement.
Upload source files in bulk, pick the categories they belong to, and mark anything you do not want licensed.
We check originality, run dedupe and screening, then write and verify the captions, tags and attributes.
Accepted files join a versioned dataset release that AI companies licence non-exclusively.
You take a share of what each release earns, split by how many of your files it carries, paid monthly.
Non-exclusive training grants on versioned releases. Evaluate before you commit, and commission what the catalogue does not already hold.
A representative sample with the full record schema, so you can test loaders and label quality before anything is signed.
A versioned release of one or more categories, licensed for model training on an annual, non-exclusive term.
You define the taxonomy, style range, volume and edge cases. Creators draw to that brief, and the set is yours first.
Your S3 or GCS bucket, a private Hugging Face repo, or signed archive links.
WebDataset shards or Parquet, with sources and renders addressed by checksum.
Immutable release tags. Additions ship as new tags; withdrawals are listed explicitly.
Consent manifest, QA report and a provenance summary written for a model card.
No. You keep full ownership. What you grant is a non-exclusive licence for the work to be included in dataset releases used for model training. You can keep selling, printing and licensing the same files anywhere else, at the same time.
Each release earns licence revenue. A fixed share of that revenue is distributed across the creators whose files are in the release, weighted by how many of their accepted files it contains. Payouts run monthly, and every line is traceable to a release and a buyer.
Yes, for anything that has not shipped. Withdrawing removes your files from all future releases and refreshes. Releases already licensed remain valid for their term, because a buyer has already trained on them — that limit is stated plainly in the grant you sign, before you submit.
Perceptual hashing against the existing catalogue and known stock libraries, plus trademark and likeness screening and a human review pass on anything flagged. Accounts that submit work they do not own are removed and their files pulled from the next release.
Yes. Scoping goes down to the sub-collection — icon sets alone, or floor plans alone, or festival illustration alone. You are not obliged to take the whole catalogue to get the part you need.
Yes, and you should. A 191-file review sample is published for the three shipping collections — stratified rather than cherry-picked, drawn at random within each format, with nothing retouched. It ships with the per-file metadata index, the catalogue index and the licence note, so you can measure the claims on this page yourself before a contract exists.
No. No generative model produced any part of any file in the shipping packs, and the licence note that travels with each pack states it. That matters more than it used to: a design model trained on another model’s output inherits its artefacts, and there is no way to detect that after the fact from the files alone.
That is what commissioned sets are for. Bring a taxonomy and coverage targets, and the work gets drawn to your brief by creators who opted into it, with the same consent and provenance record as everything else.
Join the waitlist
Creators
Send a link to your portfolio and tell us roughly what is in your archive. We will come back with the categories that fit and what a submission looks like.
Apply as a creator →AI companies
Describe the model and the gaps in your current data. We will send the record schema and an evaluation slice from the closest categories.
Request dataset access →