Datasets:

Louis0411
/

StateWorldModel

license: other
license_name: flux-2-dev-non-commercial-for-images
task_categories:
  - text-to-video
configs:
  - config_name: worldstate_bench
    data_files: data/cases.parquet
tags:
  - world-models
  - benchmark
  - interactive-video
  - game-generation
size_categories:
  - small

StateWorldModel — assets & benchmark dataset

Companion data for the code release of "State World Models: Execute the Rules, Generate the Pixels" (StatefulHarness + WorldState-Bench).

Paper: State World Models: Execute the Rules, Generate the Pixels (arXiv:2610.26811)

One repo, two independent subtrees — download only what you need:

Subtree	Size	What it is
`assets/`	~94 MB	First-frame images for the 20 games of StatefulHarness (`agentic_game/`): `assets/adv/<character>/` (story games), `assets/duel/` (combat), `assets/openworld/` (6 scenes) — consumed by the game code, fetched via `download_assets.py`
`data/cases/`	~63 MB	WorldState-Bench: 50 counterfactual hidden-state twin pairs (B01–B50, 100 clips) — shared FLUX.2 first frame + executable rule program + pre-written judge questions per twin
`data/cases.parquet`	4 MB	Flattened metadata table (one row per twin, 100 rows) with an embedded 640px first-frame preview per row (`preview`); the full-resolution frame is at `image_path` — the dataset-viewer table

Both subtrees mirror the layout of the code repository, so the in-repo downloaders place files where the code expects them:

# inside a checkout of the code repository
cd agentic_game    && python download_assets.py   # → agentic_game/assets/
cd worldstate_bench && python download_data.py    # → worldstate_bench/data/cases/

Mirrored on ModelScope: Louis0411/StateWorldModel (same layout; download with the modelscope SDK if that is more convenient).

Dataset structure (WorldState-Bench)

Each data/cases/B*/ folder is one counterfactual pair:

input.png — the shared first frame (FLUX.2, 1280×720, audited so that no visual cue reveals the hidden value; EXIF carries seed/steps/gate)
case_A.json / case_B.json — the two twin records: instruction, hidden state, executable rule (DSL), gold endings, and the pre-written fixed-choice judge questions. Expected outcomes are derived mechanically from the rule program (no human-written gold).
card_A.txt / card_B.txt — human/model-readable cards

Field-by-field schema, the generation pipeline, and the evaluation protocol: see README.md / USAGE.md in the worldstate_bench/ folder of the code repository.

License / terms

Code (the companion repository): Apache-2.0.
The first frames in data/cases/ (and the adv/openworld images in assets/) are generated with FLUX.2-dev; the FLUX.2 [dev] Non-Commercial License governs these derived images.
Baseline wrapper code in the repository that adapts HunyuanVideo-1.5 retains the Tencent Hunyuan community license header.

Citation

@article{stateworldmodels2026,
  title   = {State World Models: Execute the Rules, Generate the Pixels},
  author  = {Chan, Yan Ki and Zhang, Yanran and Zheng, Wenzhao and Wang, Xiaofeng and Zhu, Zheng and Chen, Lei and Zhou, Jie and Lu, Jiwen},
  journal = {arXiv preprint arXiv:2610.26811},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.26811}
}

Preprint entry; TODO: replace with the camera-ready entry on acceptance.