junchaoh-cs/SolarWM-Data_test-set-v1

SolarWM Standalone Test Set v1

This repository contains the complete, self-contained SolarWM test set without the training shards. It includes 1,300 clips from 13 source views (100 per view), packaged as 60 uncompressed WebDataset tar files totaling approximately 77.4 GB.

The test identities are the current accepted SolarWM standalone evaluation split, excluding MIND. The release contains 757 xhigh and 543 high samples. All selected samples have non-empty captions and finite camera trajectories.

Download

Download the full test set:

python -m pip install --upgrade huggingface_hub
hf download junchaoh-cs/SolarWM-Data_test-set-v1 \
  --repo-type dataset \
  --local-dir /path/to/SolarWM-Data_test-set-v1

Download one source only, for example sekai_game:

hf download junchaoh-cs/SolarWM-Data_test-set-v1 \
  --repo-type dataset \
  --include "wds/sekai_game/**" \
  --local-dir /path/to/SolarWM-Data_test-set-v1

Layout

SolarWM-Data_test-set-v1/
|-- README.md
|-- manifest.json
|-- SHA256SUMS
|-- indexes/all.jsonl.gz
|-- selections/all.jsonl.gz
|-- licenses/
|-- schemas/
`-- wds/<dataset>/
    |-- index.jsonl.gz
    |-- selection.jsonl.gz
    |-- stats.json
    `-- shards/test-NNNNNN.tar

Every sample contains exactly four members:

• *.video.mp4: source test video;
• *.camera.npz: camera trajectory;
• *.intrinsics.npy: camera intrinsics;
• *.manifest.json: portable prompt, media, camera, selection, and member metadata.

The root indexes/all.jsonl.gz covers all 1,300 samples. Each wds/<dataset>/index.jsonl.gz is a 100-row source-specific view. Index shard paths are relative to the repository root, so the downloaded tree can be moved to any local absolute path.

Use the indexes as the sample authority. Do not infer a different sample set by globbing tar files. SHA256SUMS records the byte identity of every payload and published control file.

Data contract

• Source video, camera, and intrinsics bytes are preserved; videos were not transcoded for this release.
• Embedded manifests were regenerated with a small public schema. Internal storage paths, processing logs, credentials, and production-control metadata are not included.
• Camera arrays retain their accepted source convention: authoritative c2w, no additional axis flip, and no Context-as-Memory /380 normalization.
• The current sekai_game payload uses its accepted metric-scale camera data.
• This repository intentionally omits the duplicate per-clip corpus layout and all build/control archives. The WDS payload is the complete test set.

See manifest.json for the exact release inventory and schemas/ for the public metadata contracts.

Sources and terms

This is a multi-source dataset. There is no blanket license that replaces or broadens the upstream terms. Media, annotations, and derived metadata remain subject to the applicable source entry in licenses/source-registry.json. The *-clean views remain subject to the terms of their corresponding upstream source.

Review licenses/DATA-TERMS.md and the linked upstream terms before download, use, or redistribution. In particular, some sources restrict commercial use.

Citation

If you use this test set, SolarWM-Data, or the released models, please cite:

@misc{huang2026solarwmopendatascalable,
  title={SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models},
  author={Junchao Huang and Guian Fang and Shengju Qian and Xianghao Kong and Zhuoran Zhao and Wei Huang and Yihua Du and Zixin Zhang and Justin Cui and Yuchao Gu and Yukang Chen and Xinting Hu and Tianyu He and Shaoshuai Shi and Zhuotao Tian and Xin Wang and Mike Zheng Shou and Li Jiang},
  year={2026},
  eprint={2609.02886},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.02886}
}

Project page: https://junchao-cs.github.io/SolarWM-Web/