遇见数据集

Phycos

收藏
魔搭社区2026-08-23 更新2026-08-23 收录
官方服务:

资源简介:

# PhyCoS This release contains the cleaned 253-case PhyCoS image-to-video dataset. ## Contents - `train.jsonl`: 253 trainable image-to-video rows. - `cosmos_sft_manifest.jsonl`: the native-frame/FPS official Cosmos Nano SFT manifest; its `vision_path` values resolve within this release directory. - `videos/`: silent H.264 MP4 targets. - `first_frames/`: paired conditioning frame 0 images. - `provenance/`: source revision/license/attribution records and a native timestamp audit; these files are metadata only and are not consumed by the SFT loader. - Case names are derived from the caption and physics category. All manifest media paths are relative to this directory. Source FPS, duration, and frame count are preserved per row. Captions use the fixed Verified-style three-sentence prefix followed by the scene-specific setup and action. The timing audit marks 16 rows with non-regular adjacent PTS intervals and 36 rows whose `avg_frame_rate` and `r_frame_rate` metadata disagree. Those rows remain available with their native timing, but the release does not claim strict CFR for them. See `provenance/temporal_audit_summary.json` for the reproducible thresholds and `provenance/temporal_audit.jsonl` for per-case flags. Source revision, license status, and attribution are recorded in `provenance/sources.json`.

提供机构:
maas
创建时间:
2026-08-09
二维码
社区交流群
二维码
科研交流群
商业服务