Datapoint AI releases SVG-Bench
By Steven Van ·
SVG-Bench is an open dataset of 352,270 human votes on text-to-SVG generation, with 14 models ranked across eight categories and a public leaderboard.
Datapoint AI introduced SVG-Bench on October 22, 2026, an open human-preference dataset for models that write SVG from a text prompt. The same release publishes the leaderboard those votes produce. The announcement is on X, the live board is Datapoint SVG Bench, and the files are the Hugging Face dataset.
What is in the dataset
The card is titled “Text-to-SVG human preferences: 352K votes across 14 models.” It holds 352,270 validated pairwise votes, exactly 10 for each of 35,227 SVG pairs. Fourteen language models wrote SVG for 400 original prompts, 50 in each of eight categories: icons, logos, illustrations, scenes, diagrams, typography, UI, and precision. 20,323 annotators in 183 countries took part. Every output is stored as SVG source and as the PNG the annotators saw, including generation cost and latency. The set is built on the Datapoint annotation platform.
How the comparisons were judged
Each pair was shown to 10 people, so a comparison has a margin rather than a single label. All 91 model pairings were scheduled on every prompt. A pair was dropped when one side produced no usable SVG (1,150 of 36,377). Chromium rendered each SVG at 1024×1024 pixels on white. Annotators saw the two renders and the prompt and were asked “Which image do you prefer?” Model names were not shown. The train and test splits hold out 40 prompts, five per category, so a reward model can be trained without those prompts leaking into evaluation.
Who leads the board
Elo ratings come from a Bradley-Terry fit on the raw votes, centered so the 14 models average 1000. The overall board gives each of the eight categories an equal share and matches the live leaderboard. GPT-6 Astra from OpenAI is first at 1065.6 Elo, with a 59.8% win rate. Gemini 3.8 Flash from Google is second at 1059.2. Muse Wonder from Meta is third at 1047.3. Claude Fable 5.1 from Anthropic is fourth at 1028.5. Arrow 2 Telos from QuiverAI is eighth overall at 1011.6 and leads the Illustrations and Scenes categories.
License and how to load it
Votes, prompts, and metadata are released under CC-BY-4.0 with attribution to Datapoint AI. The dataset terms also say not to try to re-identify annotators. The SVGs and their renders are outputs of the listed third-party models, distributed for research and evaluation, and their usage rights follow each provider’s terms. Datapoint says the SVGs are one output per model and prompt, collected for evaluation, and are not a curated training corpus for SVG generation. Load a split in Python with load_dataset("datapointai/text-to-svg-human-preferences-352k", "pairs", split="train"). The configs are svgs, pairs, responses, prompts, and models.
Related resources
Other public comparison boards on Creators Toolbox include the AI Image Generator Leaderboard and Arena AI. QuiverAI is the maker of Arrow 2 Telos, one of the models on this board. Ideogram introduced Ideogram Vector on October 22, its own model for editable SVGs from a text prompt or an image.