SVGFaith-Decomposer

Domain-tuned decomposer for SVGFaith. It reads a text prompt and emits a dependency graph of atomic visual requirements. It does not see the image.

The released SVGFaith-Score pairs this model with the unmodified Qwen/Qwen3.5-9B visual question answering backend. Aggregation is unique-ancestor gating with alpha 0.5.

Training follows the paper: teacher-generated supervised fine-tuning from Qwen3.5-397B-A17B, rejection-sampling supervised fine-tuning, then Group Relative Policy Optimization. These weights are the GRPO checkpoint exported in bfloat16.

Code: https://github.com/Yukinonooo/SVGFaith

Benchmark: https://huggingface.co/datasets/Yukinonooo/SVGFaith