Most 3D-generation datasets provide meshes or images; this one makes the executable program part of the training example. That shift matters because a model can be evaluated not just on how a render looks, but on whether its code runs and produces a usable 3D asset.
What Sets It Apart
- Entries pair code with captions and rendered views, connecting text or image conditions to executable 3D output.
- The corpus spans multiple coding dialects and sources, including Blender Python, CadQuery, GLSL, URDF, OpenSCAD, and three.js. This makes it useful for studying code generation across different kinds of 3D work, not just object modeling.
- The dataset includes execution and quality signals for some sources, such as shader compilation status, and keeps variants and caption types available for filtering. That lets researchers build more targeted training or evaluation subsets.
- The Hugging Face card describes more than two million text-only rows; image-bearing data and some staged source collections have separate availability. Check the specific split and source status before planning a run.
Great Fit If…
It suits researchers fine-tuning or benchmarking models that translate text or images into procedural 3D code, especially when execution matters alongside visual quality. Look elsewhere if you need a general-purpose 3D asset library, a uniformly verified corpus, or data cleared for commercial use: access is gated for non-commercial research, and source-specific licenses also apply.
