Note Training data for vectorizing academic figures into SVG: 66,579 figure — Complex-Diagrams (60,034 real arXiv figures) and Shapes-and-Arrows (6,545 procedurally generated diagrams).

Note Evaluation benchmark for VFIG with two test configs: VFIG-Bench (400 in-distribution figure–SVG pairs held out from VFIG-Data) and VFIG-Bench-OOD (198 image-only out-of-distribution academic figures for generalization probing).

XunmeiLiu/VFIG-4B

Reinforcement Learning • 4B • Updated Mar 27 • 217 • 5

Note A 4B open-source vision-language model for complex figure-to-SVG conversion, trained with curriculum SFT + RL refinement.