Rapidata SVG · Human Preference
Rapidata · Humans compare vector graphics generated by models and directly rate visual preference.
What it measures, and how
Only the Preference Elo of the same static SVG task is used; Alignment, Coherence and Overall are not counted twice. The public score table of the fixed version of the official dataset is the basis, and the original human battles can be traced.
How this evidence is used
Under observation: official data and licence are available, but three DeepSeek rolling aliases lack fixed-version evidence, so the whole score is not adopted for now. Awaiting the evaluator's run versions; the budget for active sources has not been cut.
Limits and data attribution
SVG drawing cannot represent all interface design. The public raw data contains no model version, generation date or API fingerprint, so it is impossible to tell whether the rolling interface has switched versions; the leaderboard update date cannot be used to guess the model version.
Data licence: CC BY 4.0 · Rapidata/svg-benchmark; retain official attribution and data links.
Scores published by Rapidata. Raw scores and the News consensus score use different scales and cannot be added directly.