SolarWM-Data

# 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 and citations 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. SolarWM paper citation and BibTeX entry: **Coming soon.**