README.md

---
pretty_name: SolarWM-Data
license: apache-2.0
language:
  - en
tags:
  - video
  - webdataset
  - world-model
size_categories:
  - "1M<n<10M"
---

# SolarWM-Data

SolarWM-Data is a reusable video-data foundation for camera-conditioned
world-model research. The main Hugging Face repository publishes portable
release controls, licenses, deterministic test indexes, and directly readable
format examples. It also contains the `SolarWM-Data-Annotation/`
reconstruction package. The full raw video and preencoded latent payloads are
distributed separately because of their size and upstream terms.

Project Page: [SolarWM](https://junchao-cs.github.io/SolarWM-Web/)

```text
SolarWM-Data/
  README.md
  SolarWM-Data-Annotation/
  releases-v1/
    release.json
    checksums.jsonl.gz
    recipes/
      clean-81f/
      clean-153f/
      clean-158f-h3/
      clean-957f/
    test-set/
    example/
    licenses/
    raw-wds/      # obtained separately when a raw example needs it
    latent-wds/   # separately published, backend-specific generations
```

`release.json` and `checksums.jsonl.gz` describe the complete logical release
across its distribution repositories. Their presence in the main repository
does not mean that all raw and latent payloads are stored in that repository.
Recipe rows use paths relative to `releases-v1/`, so separately downloaded
payloads can be placed into the same local tree.

## Download the main repository

```bash
python -m pip install --upgrade huggingface_hub
hf download junchaoh-cs/SolarWM-Data \
  --repo-type dataset \
  --exclude "SolarWM-Data-Annotation/**" \
  --local-dir /path/to/SolarWM-Data
```

The small `releases-v1/example/` archives support schema inspection and reader
or preencoding smokes. They are not a training corpus. Remove the `--exclude`
option to download the Annotation package too, or use `--include
"SolarWM-Data-Annotation/**"` to download only that package.

## Add the payload needed by your run

There are three routes:

1. **Reconstruct raw-WDS from annotations.** The
   [`SolarWM-Data-Annotation/`](SolarWM-Data-Annotation/README.md) directory in
   this repository contains annotations, public source identities, and the
   reconstruction tools. Its three `*-clean` owners also include the processed
   videos, so they restore without a model or GPU. For the other 11 owners,
   users acquire the original videos under their respective terms and follow
   that directory's README to rebuild `raw-wds/`.
2. **Request raw-WDS access.** Use the Raw-WDS Access entry on the
   [SolarWM project page](https://junchao-cs.github.io/SolarWM-Web/).
3. **Download preencoded data.** Every Wan, LTX, and MiniMax-H3 latent
   generation will have a separate dataset repository. Links will be added to
   the SOLAR-WM code repository's `docs/data-access.md` as uploads complete.

Preserve raw payloads under `releases-v1/raw-wds/` and latent payloads under
`releases-v1/latent-wds/<generation>/`. The selected SOLAR-WM example's
`train_index`, `index`, and `test_index` fields state exactly which payloads it
needs.

## Complete annotated corpus

The 14-source raw corpus contains 1,425,694 samples: 471,708 high, 404,545
xhigh, and 549,441 rejected. Rejected shards are part of the release. Every
sample retains `kept`, `kept_tier`, `reject_reasons`, and the available camera,
motion, quality, scene, and VLM measurements.

Source preprocessing and training-mixture construction are separate. Users can
change thresholds, tier policies, sampling ratios, and source weights without
rerunning video decoding, camera estimation, VMAF, UniMatch, DOVER,
saturation, scene-cut, or VLM processing.

The source directories are `abot`, `dl3dv-10s`, `dl3dv-60s`, `mind`,
`miradata`, `miradata-clean`, `multicamvideo`, `omniworld`, `realcam_vid`,
`sekai_game`, `sekai_walking`, `sekai_walking-clean`, `spatialvid`, and
`spatialvid-clean`. Each directory contains tiered WebDataset shards,
`meta.jsonl`, and complete tier/sample indexes.

## Preencoded data

`latent-wds/` provides reader-ready generations for Wan 2.2 TI2V-5B at
81f, 153f, and 957f with published 480P/720P variants; Wan 2.2 I2V-A14B at
81f, 153f, and 957f with the
published 480P/720P variants; MiniMax-H3 at 158f/768P; and LTX-2.5 video-only
at 153f and 953f. Tensor member manifests declare their `solarwm_*` schema,
source identity, shape, dtype, and camera contract.

Camera trajectories and intrinsics are included in each record and are ready
for the corresponding SolarWM reader.

## Recipes and evaluation

Recipe indexes use relative object keys so the same controls work with a local
copy or with bucket streaming. Evaluation selects a deterministic subset from
the matching recipe `test-index.jsonl.gz` using `sample_count` and
`selection_seed`.

`test-set/` is a logical view of canonical primary raw shards. It preserves
selection rank, split identity, and minimum-frame requirements without storing
a second copy of each video.

`example/` contains one raw representative from each annotation tier and one
representative of every declared release latent generation. Each example record names
the release object from which it was derived.

## Local use

Set both SOLAR-WM paths to the same local release root. A local copy may live
at any absolute path:

```yaml
data:
  index_root: /path/to/SolarWM-Data/releases-v1
  transport:
    kind: local
    root: /path/to/SolarWM-Data/releases-v1
```

Indexes resolve shard paths relative to that root. The checksum catalog is
available when a complete release-integrity check is needed.

## Licenses

Media and annotations retain their upstream terms. There is no blanket license
that replaces source-dataset restrictions. Read
`releases-v1/licenses/source-registry.json` and the linked upstream terms before
use, redistribution, or creation of derived data.

## Citation

**If you use SolarWM-Data, the data engine, or the released models in your research,
please cite our paper.**

Paper: https://arxiv.org/abs/2609.02886

```bibtex
@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}, 
}
```