SpaceCode-Bench 1.0.0-dev

SpaceCode-Bench is a procedural development benchmark for constructing, reconstructing and editing 3D environments with executable Blender code.

This release pairs the frozen 1.0-procedural collection with evaluator 7: 1,200 native scenes, 3,600 task bundles and 4,800 execution conditions across 40 profiles and 240 lineages. The dataset includes full prompts and contracts, native Blender assets, calibrated observations, exact integrity manifests and a complete authored edit example.

The evaluator enforces fixed track dimensions, leaves incomplete required measurements unscored, rejects inconsistent imported results and runs numerical scoring in a bounded process with verified dependencies. Results expose coverage, critical failures, dimensions, family aggregates and lineage uncertainty. The project website provides interactive examples, dataset statistics, the evaluation procedure and local run comparison.

The source archive includes the hydrated website assets for offline use. The Linux amd64 Blender 4.2.23 runtime archive retains its original runtime validation; publishing it does not claim a new evaluator-7 Docker run. Verify the accompanying SHA256 files before extraction or loading. Native data is hosted on Hugging Face.

This is a development release. Semantic identity (C), visual finish (V) and full_score remain unmeasured; there are no formal target-model rankings or human calibration claims. See the repository's native protocol, data card and independent automated review records for scope and limitations.

Project website · Dataset · Documentation

Validation includes 427 Python tests, 49 frontend tests, both Python CI versions, independent automated reviews and actual public-service readback. The complete source/runtime attachments and Hub files were anonymously read and checked against SHA256. Verification record · Supplementary review records.