Text2CAD-GC

Text2CAD-GC is a geometric-continuity benchmark and reinforcement-learning toolkit for text-to-CAD, graded on G0, G1 and G2 continuity. It tests a question that global Chamfer distance, rendered images and watertightness hide: do adjacent generated surfaces actually achieve positional continuity (G0), tangent-plane continuity (G1) or curvature continuity (G2) across their shared seam? A generated model can have a low point-cloud distance and still mesh while adjacent NURBS faces lack shared topology, contain positional gaps, show normal creases or exhibit curvature jumps.

The current release is a reproducible pilot rather than a paper-scale benchmark: 90 analytic two-patch STEP models with exact labels (30 each for G0, G1 and G2), a fixed 60/15/15 train, validation and test split, an Open CASCADE seam-level continuity evaluator, a shaped reward for PPO, GRPO or verifier-guided reinforcement learning, 3,691 de-identified derived seam measurements from a PartABC audit without redistributing restricted source CAD, and guidance for combining continuity evaluation with existing text-to-CAD benchmarks. The README is bilingual, Chinese and English.