README.md

---
license: apache-2.0
language: [en, ja]
task_categories: [text-generation]
tags: [cad, openscad, cadquery, 3d, code-generation, synthetic]
pretty_name: OpenSCAD/CadQuery Spatial Code v0 (pilot)
size_categories: [n<1K]
configs:
- config_name: default
  data_files: [{split: train, path: data/train.jsonl}, {split: test, path: data/test.jsonl}, {split: hold, path: data/hold.jsonl}]
---
# spatial-code v0 (pilot)
Small, fully synthetic text→CAD-code pairs (OpenSCAD and CadQuery) for simple parametric solids: boxes, cylinders, spheres, tubes, plates with 4 holes, cube−sphere differences, cube+cylinder unions, L-brackets.

## Construction
- Template generator, `random.Random(42)`; code and prompts (EN + templated JA) generated from the same parameters.
- Every row executed: OpenSCAD 2021.01 CLI → STL, CadQuery 2.8.0 → STL; mesh checked with trimesh (non-empty, finite bounds, watertight, volume>0). Failures excluded (see QUALITY.md).
- Exact-code dedupe. Split by hashed prompt (80/10/10 target; realized 147/13/18 over 178 rows).

## License
Apache-2.0. All content self-generated; no third-party code or models used for content.

## Limitations (please read)
- Pilot scale (178 rows, 8 shape families); low diversity, mostly numeric variation. Not a benchmark.
- Prompt↔geometry correctness is by construction, not independently measured.
- CadQuery spheres/cube−sphere were rejected by the watertight check, so dialect coverage is uneven.
- Japanese prompts are template translations, not native-written.
- Small test/hold splits; statistics on them are high-variance.