---
license: cc-by-4.0
language:
- en
- zh
pretty_name: SpaceCode-Bench
size_categories:
- 1K<n<10K
task_categories:
- text-generation
tags:
- benchmark
- blender
- code-generation
- 3d-environments
- procedural
configs:
- config_name: default
  data_files:
  - split: train
    path: catalog/train/*.parquet
  - split: validation
    path: catalog/validation/*.parquet
  - split: test
    path: catalog/test/*.parquet
---

# SpaceCode-Bench

**Construct, reconstruct and edit native 3D environments with Blender code.**

[Website](https://kurtdu.github.io/SpaceCode-Bench/) · [Code and evaluator](https://github.com/KurtDu/SpaceCode-Bench) · [Protocol](https://github.com/KurtDu/SpaceCode-Bench/blob/main/docs/NATIVE_PROTOCOL.md) · [Complete worked example](https://huggingface.co/datasets/KurtDu/SpaceCode-Bench/tree/main/examples/reading_library__open_hall__00)

SpaceCode-Bench is a procedural development benchmark with **1,200 materialized scenes**, **3,600 task bundles** and **4,800 execution conditions**. Each task publishes a complete numerical contract. Evaluation reopens the native `.blend`, extracts geometry and controlled views, and measures the declared requirements.

Construction starts from text; reconstruction uses one or four calibrated images; editing starts from a supplied native scene. The 40 space profiles cover indoor and outdoor environments across six topologies and five geometric variants. The shared asset grammar has 46 furniture and facility categories.

## Catalogue and native data

The dataset viewer has one row per execution condition. `prompt` contains the full solver request, `contract_json` the complete public contract, and `contract_sha256` its identity. `task_path` locates the task inside the archive named by `archive_file`. `archive_file` is relative to the Hugging Face repository root; `task_path` is relative to the extracted data root. `base_environment_preview` shows the authored base environment for discovery. It is neither an additional solver input nor the edited target; edit starters and targets are in the native archives and website walkthrough.

```python
from datasets import load_dataset

tasks = load_dataset("KurtDu/SpaceCode-Bench")
example = tasks["train"][0]
print(example["task_id"])
print(example["prompt"])
```

The Parquet catalogue supports discovery. The 12 archives in `archives/` contain the complete executable data: original native scenes, original PNG observations, public contracts, starters, private evaluator mappings and per-reference validation evidence. `dataset-release.json` identifies every archive and checksum. The GitHub downloader pins an immutable Hub revision and verifies the assembled contents:

```bash
git clone https://github.com/KurtDu/SpaceCode-Bench.git
cd SpaceCode-Bench
python -m venv .venv
source .venv/bin/activate
python -m pip install -c requirements-tested.txt .
python scripts/download_dataset.py
python scripts/verify_release.py --root data/SpaceCode-Bench-1.0.0-dev-data
```

The compressed data occupy 3,146,764,769 bytes; the complete assembly occupies 5,530,090,296 bytes. Mount only the resources specified by a task into a solver, never the complete data root. The [worked example](https://huggingface.co/datasets/KurtDu/SpaceCode-Bench/tree/main/examples/reading_library__open_hall__00) provides a full public prompt, native starter and edited scene, builder, independently extracted evidence and clause-level scoring.

## Splits and units

| Split | Scenes | Conditions |
| --- | ---: | ---: |
| Train | 895 | 3,580 |
| Validation | 110 | 440 |
| Test | 195 | 780 |

A lineage is a profile/topology pair. All five variants and all task conditions of each of the 240 lineages remain in one split. These are same-generator partitions with public references. They do not provide a hidden test set, contamination resistance or demonstrated out-of-distribution transfer.

Coordinates use metres and Z up. Reconstruction supplies exactly one or four matching camera/image pairs at 480×320 pixels. Difficulty labels describe authored topology classes, not measured model difficulty. The collection has six edit operations with 200 scenes each, eight prompt forms with 150 scenes each, 900 English scenes and 300 scenes with Chinese intent plus English numerical requirements.

## Evaluation scope

Native dimensions measure architecture/floors (A), geometric regions and negative space (G), prescribed route clearance/support (F), supplied-view agreement (R) and protected edit geometry (P). The track sets are AGF for construction, AFR for reconstruction and AGFP for editing. The score is their geometric mean scaled to 100; a measured critical failure caps it at 49. Required missing or unresolved measurements produce no aggregate score, even if another failure is already known.

Semantic identity C, visual finish V and full_score remain null. Geometric spans do not establish semantic categories or counts. Protected geometry does not certify material preservation. Physical stability, arbitrary navigability, photorealistic quality and hidden-view correctness are outside this native protocol. Observation thresholds are public development settings, not human-calibrated accuracy estimates.

No formal model results are included. Author controls test the evaluator; they are not model baselines. Results should retain the full planned population, report coverage and failures, distinguish single/multi-view conditions, and estimate uncertainty by resampling lineages within profile families. The [protocol](https://github.com/KurtDu/SpaceCode-Bench/blob/main/docs/NATIVE_PROTOCOL.md) specifies numerical limits and reporting rules.

## Provenance and version

Dataset version: `1.0-procedural`. Protocol: `1.0-native`.

Dataset fingerprint: `6791f24b18c04f8a9b0aa3d8a003ac52147c360a002d6bce70930689afc0448c`.

The frozen corpus records its evaluator-6 author admission and original generation/extraction identities. The current evaluator-7 adds stricter score admission, required-measurement handling, bounded numerical execution and actual dependency checks. New checks are recorded separately; historical evidence is not relabeled. Archive SHA256 manifests bind all payload files and the full data collection.

Geometry is authored with the repository's procedural grammar. No source photographs, decoration archives or commercial assets are distributed. Data are licensed under CC BY 4.0; accompanying code under MIT.

## Citation

Yuhao Du. *SpaceCode-Bench*, version 1.0-procedural, 2026. Record the dataset fingerprint, Hub revision and evaluator fingerprint used in an experiment. Citation metadata is provided in [CITATION.cff](CITATION.cff).