Hugging Face's logo

Datasets:
shivamk3r
/
shape-code-bench-eval-v1

Modalities:
Image
Text
Formats:
parquet
Size:
< 1K
ArXiv:
Tags:
image
Synthetic
benchmark
code-generation
program-synthesis
arxiv:2605.11680
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
shape-code-bench-eval-v1 / README.md
shivamk3r's picture
shivamk3r
Add ShapeCodeBench dataset card
4839b59 verified
preview code
|
Raw
History Blame Contribute Delete
2.23 kB
metadata 
pretty_name: ShapeCodeBench eval_v1
license: cc-by-4.0
size_categories:
  - n<1K
tags:
  - image
  - synthetic
  - benchmark
  - code-generation
  - program-synthesis
  - arxiv:2605.11680

ShapeCodeBench eval_v1

This dataset is the frozen eval_v1 reporting split for ShapeCodeBench, a synthetic benchmark for testing whether multimodal models can reconstruct executable drawing programs from rendered shape images.

It contains 150 grayscale 512x512 PNG images: 50 easy, 50 medium, and 50 hard examples. Each row includes the rendered image, the canonical ShapeCodeBench DSL program that generated it, generation metadata, render configuration, and the SHA256 checksum for the source PNG.

Zenodo remains the archival release DOI: https://doi.org/10.5281/zenodo.20132286. This Hugging Face dataset is a discoverable and loadable mirror of the frozen evaluation split.

Load 

from datasets import load_dataset

dataset = load_dataset("shivamk3r/shape-code-bench-eval-v1")

Columns

  • image: rendered target image.
  • image_file_name: original path under data/eval_v1.
  • image_sha256: checksum from SHA256SUMS.
  • sample_id, split, difficulty, seed: sample identity and generation seed.
  • image_size, num_shapes, shape_inventory: scene summary metadata.
  • ground_truth_program: canonical ShapeCodeBench DSL program.
  • render_config: deterministic V1 renderer configuration.

Evaluation Hygiene

eval_v1 is a frozen reporting split. Do not tune prompts, adapters, model checkpoints, heuristic parameters, or generator settings on this split and then report the result as clean held-out performance.

For development, generate separate train/dev splits from fresh seeds using the ShapeCodeBench repository.

Links

License

The generated benchmark dataset is licensed under CC BY 4.0. ShapeCodeBench source code is licensed separately under MIT.