Datasets:

Languages:
English
ArXiv:
License:
RuleVerse / README.md
zmling's picture
Document RuleVerse split layout and in-progress release
1f4dc73 verified
|
Raw History Blame Contribute Delete
2.4 kB
metadata
license: apache-2.0
language:
  - en
pretty_name: RuleVerse

RuleVerse

RuleVerse is a paired dataset for WorldRuler: Interactive World Modeling under Configurable Rules. It contains rule-enabled and rule-disabled videos from Habitat and GTA-V, with rule definitions and annotations of state and execution.

Project page · Paper · Code

Release status

Upload in progress. Consult release-manifest.json for the list of successfully uploaded archives and their SHA-256 checksums. A split is incomplete until all its archives are available; the manifest will report complete when the full release has been uploaded.

Dataset size

Environment Split Pairs Videos
GTA-V Train 900 1,800
GTA-V Test 100 200
Habitat Train 2,000 4,000
Habitat Test 200 400
Total 3,200 6,400

The release includes the original videos and annotations from the training and test splits. Reserve candidates are excluded. Videos are not re-encoded.

Download and extract

hf download AlayaLab/RuleVerse --repo-type dataset --local-dir RuleVerse

To download only the GTA-V test split:

hf download AlayaLab/RuleVerse --repo-type dataset --include 'gtav/test/*' --local-dir RuleVerse

Each split contains metadata.tar.gz and numbered pairs-*.tar.gz archives. Extract all downloaded archives into the same output directory; each archive already includes the environment and split in its paths:

mkdir -p RuleVerse-extracted
find RuleVerse -name '*.tar.gz' -exec tar -xzf '{}' -C RuleVerse-extracted \;

The restored layout is gtav/{train,test}/ and habitat/{train,test}/, with each split's original index.json, pairs/, and supporting metadata. Use the split index to locate episode videos and annotations. Each pair contains a positive (rule-enabled) and negative (rule-disabled) episode.

Citation

@article{zhai2026worldruler,
  title={WorldRuler: Interactive World Modeling under Configurable Rules},
  author={Mingliang Zhai and Chuanhao Li and Jiaming Tan and Zhen Li and Yuwei Wu and Yunde Jia and Zian Meng and Yuanyang Yin and Kaipeng Zhang},
  journal={arXiv preprint arXiv:2611.24869},
  year={2026}
}