A rendered image can show that a scene looks plausible, but it cannot by itself establish that the code behaves correctly. This dataset pairs code with visual references and separates related concepts in validation, making it useful for studying both generation quality and generalization—while leaving animation and interaction to be checked across multiple states.
What Sets It Apart
- The 3,445 examples span 50 domains, from physics and biology to history, games, and everyday environments. This breadth lets researchers examine whether scene-generation behavior transfers across subject matter.
- Each example pairs a prompt and assistant response with a curated reasoning trace; the response contains HTML/JavaScript. A separate preview configuration aligns rendered references by identifier, enabling visual inspection and image-conditioned experiments.
- The 182-example validation split holds out a concept family in each domain, keeping related examples together. This is a more informative test of generalization than a random split, though the validation set is comparatively small.
- All examples have rendered evidence and visual review, but a still image cannot verify every animation or interactive behavior. Treat the previews as visual references, not a complete execution test.
Who It’s For
A good fit for researchers training or evaluating models that generate Three.js scenes, especially when they need varied subject matter and paired visual references. Look elsewhere if you need a large-scale benchmark, exhaustive browser-behavior testing, or a dataset whose rights are broader than the publisher’s stated MIT license scope.