collyer12/opensora-planning-h5p5

Open-Sora H5/P5

A video dataset containing pairs of five consecutive RGB history frames and five consecutive RGB planning frames, English captions, timestamps, and source metadata. Videos are stored as deduplicated MP4 files. Parquet rows reference their exact zero-based frame indices; individual frame images and encoded latents are not included.

Available data

Item	Value
Samples	25176
Distinct source videos	4319
Updated (UTC)	2026-10-20T02:15:49.386216+00:00

Configurations: all contains all samples; current_v5 and legacy_v2 contain the corresponding annotation versions. Use all unless a specific version is needed.

Sample fields

• history_indices: five consecutive source-frame indices.
• planning_indices: five consecutive later source-frame indices.
• caption: a short English visual annotation generated by a model.
• delta_t: the time from the last history frame to the first planning frame, calculated from the frame indices and the exact rational frame rate.
• Source identifiers, video SHA256, decoded RGB hashes, frame rate, and timing metadata support verification and reconstruction. audit_json contains annotation evidence and available upstream provenance; model annotations are not human ground truth.

History precedes planning. Event intervals represent sampled windows and should not be interpreted as exact action-onset labels. No caption expansion is applied.

Sources

Source videos originate from Open-Sora-Plan-v1.1.0 / Pexels and MiraData. Available upstream revision, clip identifier, source URL, and time range are recorded in the sample metadata. Shared platform/video identifiers support deduplication across sources. Source material remains subject to its original terms; this dataset does not confer additional rights to the underlying videos.

Loading

from datasets import load_dataset

ds = load_dataset("collyer12/opensora-planning-h5p5", name="all", split="train")

Install requirements.txt and download reader.py from this repository to reconstruct RGB frames:

from reader import load_hub

pair = load_hub("collyer12/opensora-planning-h5p5", index=0, config="all")
assert len(pair["history"]) == len(pair["planning"]) == 5
print(pair["caption"], pair["delta_t"])

The reader pins metadata and video to the same repository revision and verifies video and RGB hashes. Use the decoder versions in requirements.txt for reproducible RGB reconstruction. Loading metadata alone does not download every video. For a local snapshot, use load_local(path, index=0, config="all").

Pin a repository commit for reproducible use. If creating train/validation splits, group by source-video identifier to avoid overlap. snapshot.json, batches/*.json, and CHECKSUMS.sha256 provide dataset counts, publication receipts, and file checksums.