SVGFaith: A Two-Layer Meta-Evaluation of Faithfulness Metrics for Text-to-SVG Generation
Paper • 2611.06048 • Published
Benchmark data for SVGFaith: A Two-Layer Meta-Evaluation of Faithfulness Metrics for Text-to-SVG Generation.
Code: https://github.com/Yukinonooo/SVGFaith
Decomposer: https://huggingface.co/Yukinonooo/SVGFaith-Decomposer
| Path | What it is |
|---|---|
diagnostic/ |
1,842 controlled contrastive pairs. 921 compare two renderings against one prompt. 921 compare two descriptions against one rendering. Eight capability axes. |
arena/ |
1,327 human comparisons of real text-to-SVG outputs. image_a is the more faithful rendering. |
diagnostic/manifest.jsonl is the diagnostic index. Arena pairs are arena/pairs.jsonl, with renderings under arena/images/.
Annotator identities and raw model traces are not included. Arena excludes majority ties and the 25 pairs removed in the published label review.
CC BY 4.0.