Why every public SVG dataset is icons, and what that costs a model

The shape of what is available

The public vector corpora are large and genuinely useful, and they are almost entirely icons, emoji and single-object illustrations. SVG-Stack, MMSVG-2M, SVGX-SFT-1M and SAgoge all sit in that space. The InternSVG paper puts it plainly, noting that existing sets pay insufficient attention to SVGs with specific applications in professional domains.

That is not a criticism of those datasets. They were assembled from what the open web offers, and the open web offers icons.

What an icon cannot teach

An icon is one object on an empty ground. A designed page is a system. A model trained only on icons has never seen:

• A reading order, because an icon has no second element.
• Type hierarchy, because an icon has no headline, subhead or caption.
• A grid, margins or a page format, because an icon has no page.
• A chart with a legend that names its series, because an icon carries no data.
• Text written in the context of the graphic beside it, which is the relationship a multimodal model most needs to learn.

The licence problem underneath

Scale is only half of it. Most public vector corpora are assembled from sources whose terms were never written with model training in mind, and several are explicitly non-commercial. A lab that trains a commercial model on them inherits a provenance question it cannot answer later, when a customer or a regulator asks.

That question is harder to fix than a data gap. You can collect more data; you cannot retroactively license what you have already trained on.

What full pages add

A corpus of complete pages gives a model the relationships an icon set has no way to express: where a headline sits relative to the chart it introduces, how a legend maps to the marks it labels, what a page looks like when it is dense and what it looks like when it breathes.