Many code-generation datasets stop at source code, leaving it unclear whether the output actually renders or produces a usable asset. Here, each example connects a request to executable Three.js, a rendered reference, and an exported GLB—making the dataset useful for studying the full path from prompt to inspectable 3D result.
What Makes It Useful
- The 3,918 examples span 40 domains, from characters and props to machinery and modular architecture, giving training and evaluation a wider range than a single asset category.
- Matching conversations, previews, and meshes share identifiers and splits. This makes it possible to compare generated code with both a visual target and an exported artifact.
- The examples include static, animated, and interactive assets. That variety helps explore procedural asset generation beyond single still models, though animation and interaction cannot be assessed from a preview image alone.
Best Fit and Trade-offs
Great fit if you are training or evaluating models that turn natural-language specifications into Three.js-based game assets, or building tooling around code-to-GLB workflows. Look elsewhere if you need scanned real-world geometry, broad coverage of 3D modeling tools, or a benchmark of visual quality independent of code generation: this dataset is focused on generated Three.js assets, and its examples were validated for rendering and export rather than presented as a universal quality ranking.
