About SVG Arena
SVG Arena benchmarks how well AI models illustrate — not with words, but with vector graphics. Models are given the same prompt and asked to produce a complete, self-contained SVG illustration. You compare two results side by side, blind, and pick the better one.
How it works
- Each prompt is sent to several competitors — a competitor is a model paired with a generation technique (zero-shot, best-of-N, or self-critique).
- Outputs are sanitized and stored. On the Arena, two are shown blind; you vote which is better (or tie / both bad), optionally tagging why.
- Every vote is stored as an immutable row. Together they form an open human-preference dataset; the leaderboard's Elo ratings are a byproduct.
Quality & trust
Voting is anonymous. To keep the data honest, a small fraction of matchups are attention checks(a real illustration vs. a deliberately broken control); raters who fail them are flagged and down-weighted. Votes are rate-limited per session. All model-generated SVG is sanitized server-side before it's ever rendered.
The dataset
The full preference dataset is downloadable as JSONL — one judgment per line, each with the prompt, both competitors, the winner, reason tags, and the rater's quality score.