DiagramSketch

This repository contains the full dataset archive and a smaller sample archive. The sample is provided so users can inspect the directory structure, test data loading code, and review the metadata format before downloading the full dataset.

Files

• dataset.tar.gz: full dataset archive.
• dataset_sample.tar.gz: small sample archive hosted alongside the full archive.
• dataset_sample_manifest.jsonl: lightweight manifest used by the Hugging Face Dataset Viewer to preview the sample structure.
• README_dataset_sample.md: description of how the sample was created.
• create_dataset_sample.py: script used to recreate the sample from the uncompressed dataset.
• SHA256SUMS: SHA256 checksums for the release archives.

Dataset Structure

Both the full archive and sample archive use the same top-level layout:

aesthetic_final_decision_v2.jsonl
sketches_topology_handdrawn_dgbs/
sketches_topology_handdrawn_fgbs/

The sample archive additionally includes:

README_SAMPLE.md
sample_manifest.jsonl

Sample Creation

The sample was created deterministically from the uncompressed dataset/ directory with random seed 20260507. PNG files were grouped by original figure id by removing the final _<variant>_gpt_loose.png suffix. The script sampled 30 DGBS groups and 20 FGBS groups, copied all existing variants for each selected group, and filtered aesthetic_final_decision_v2.jsonl to metadata rows matching the sampled DGBS source figures.

Recreate the sample with:

python create_dataset_sample.py \
  --dataset dataset \
  --output dataset_sample \
  --release-dir dataset_release \
  --seed 20260507 \
  --dgbs-groups 30 \
  --fgbs-groups 20

License

License information is provided in the repository metadata. Please see the dataset card and accompanying files before using or redistributing the dataset.