A 3D code dataset is more useful when it preserves the whole path from request to rendered result—not just a code snippet. These examples pair scene prompts with full HTML/JavaScript responses and generated reasoning traces, while a separate preview configuration provides rendered references. That makes the collection useful for studying both what a model was asked to build and what its code is intended to produce.
What Sets It Apart
- A sizeable, varied corpus: 4,146 scenes cover objects, interiors, architecture, landscapes, animated mechanisms, and voxel environments. This breadth helps expose code-generation systems to more than one visual style or subject type.
- Training data with a preserved validation split: 3,933 training examples and 213 validation examples make the original split available for experiments; complexity levels from 1 to 10 are annotations, not benchmark scores.
- Text, code, and visual references are separable: The default configuration holds conversation data, while optional previews align rendered images by scene ID. This supports workflows that need code alone as well as those that compare generated code with a visual target.
- Useful provenance, with a caveat: The prompts, reasoning traces, and responses were generated with MiniMax M3.1 Flash. The traces are part of the dataset, but should be treated as generated artifacts rather than verified records of a model’s internal reasoning.
Who It’s For
A good fit for researchers training or evaluating models that translate natural-language instructions into interactive Three.js scenes, especially when they need paired code and visual references. Look elsewhere for a general-purpose JavaScript corpus, real-world human-authored scene projects, or a benchmark with independently validated quality scores: this collection is focused on generated Three.js examples, and its complexity labels are not performance measurements.
