Nicholas0228/hyworld-memory-scale16-prefixes

HY memory scale16 starting images and captions

35,000 training pairs and 500 source-video-disjoint held-out pairs, curated for fresh HY-WorldPlay memory-compressor training. Images are 832×480 RGB JPEGs with image-specific English captions. The image payload is 5.96 GiB.

This initial release contains starting images, captions, provenance, split/quality reports and fixed episode/action assignments. It does not contain student-generated training futures. The 16-H100 operator prepares teacher20P prefixes and native360P held-out references with the supplied code and can publish those under prepared/ in a later dataset revision. Keep the immutable image revision pinned.

Use with the runnable pipeline

Code and complete setup/training/evaluation instructions: scale16 operator handoff.

# After installing the repository's scale16 environment:
DATA_REVISION=$(python -c 'import json; print(json.load(open("configs/hyworld_scale16/data.json"))["revision"])')
python -m hyrcm.scale_transfer fetch-images --root /shared/hyworld-scale16/data --revision "$DATA_REVISION"

The downloader verifies every tar shard and extracted image, supports interrupted downloads and repairs missing/corrupt extracted files. It removes downloaded tar files after extraction. image_shards/ contains ordinary tar files with paths matching the JSONL manifests; a custom data service is not required.

Files

• train.jsonl, validation.jsonl, manifest.jsonl: per-image metadata, captions, source IDs/timestamps, split groups, image SHA256, quality metrics and caption-model provenance.
• image_shards/, package.json: image bytes and archive SHA256/size/count manifest.
• episodes.jsonl, episodes.summary.json: full horizon/motion-family/seed assignments made before generating teacher prefixes. Training futures remain on-policy.
• reports/: extraction, deduplication, exclusions, split/integrity counts and visual-review findings.
• licenses/: retained source license and pinned license provenance.
• previews/: a deterministic stratified visual-audit sample and review, not a manual review of all 35,500 pairs.

The image manifest SHA256 is 468079434ca1a3d578ff64aeeb38092f05e48e3878f150009dca05d442f02549. The episode-plan SHA256 is fca1d6d9e32ac227cfb0a2552786aeea1277cf90e9014729a8c746ae7e80e7c3.

Sources and split

All images come from verified Sekai-Real clip/subclip IDs, matched against official Lixsp11/Sekai metadata. Distribution mirrors are zemengzero/Sekai-Project and yssssjtu/Sekai; exact pinned revisions and source timestamps are retained per record. DL3DV access was gated for the connected account and no DL3DV material is included.

The final split has 1,636 training and 64 held-out source/known-duplicate groups, with zero overlap. Exact image duplicates and conservatively confirmed perceptual duplicates were removed before splitting. Multiple distinct views from a source remain: 35,500 pairs do not imply 35,500 independent physical environments. Different source videos can still show the same location.

Image-level scene distribution

Label	Pairs
indoor	2,975
mixed	1,993
natural	2,703
other	54
urban	27,775

These are model-assigned labels for the actual starting image. Original source clip-level labels are also retained but can disagree with the starting frame. The pool remains urban-heavy; balancing cannot manufacture absent scenes.

Curation and limitations

We extract only first frames, apply conservative resolution/crop/exposure/detail checks, deduplicate, reserve source groups, caption images and screen overlays. Captions use deterministic Qwen2.5-VL-7B-Instruct at revision cc594898137f460bfe9f0759e9844b3ce807cfb5. Every selected image passed a full JPEG decode and SHA256 check. A 92-pair stratified contact-sheet sample was visually inspected.

Captions/scene labels can contain errors or inferred names; no manual correctness guarantee is made for every pair. Small watermarks may remain. The conservative overlay screen can reject usable images with ordinary signs or small watermarks, and cannot guarantee finding every overlay. Confirmed source-level exclusions and audit findings are recorded. See the curation protocol for exact rules and reproduction modules.

Motion assignments use 50% released WorldCompass sampled actions, 25% bounded outward/return paths and 25% loops/revisits. They are synthetic controls for HY generation, not recovered ground-truth camera trajectories of these source images. Held-out assignments extend to 360 latent frames (about 59.9 seconds); this is an internal evaluation protocol, not a claim of an official common external benchmark.

License

Noncommercial research use only, under the included Sekai license. Its original conditions and disclaimer are retained. Source attribution and original video IDs are preserved. Base-model licenses apply separately when preparing or training HY. This curation does not expand the rights granted by source or model licensors.

Preview

Additional sheets and the associated records are in previews/.