This repository contains model-generated annotations for a frozen selection of 40,000 videos from OpenVid-1M. It does not contain the raw videos. The source dataset is pinned to commit d8a63bd22989c80b5734ec2bb989f4e1b61a5807; the selection and download metadata are in source-index/ and release.json.

The eight ZIP files in archives/ each contain 5,000 canonical annotations/<video_id>.json.gz files. Each annotation includes entity, tracking, quality, and mask information. Masks use inline RLE, so the ZIPs do not require separate mask assets. These are model outputs, not human ground truth. Read the status field before using an annotation for training.

ZIP	Successful	Terminal failures
`set-00.zip`	4,997	3
`set-01.zip`	4,992	8
`set-02.zip`	4,997	3
`set-03.zip`	4,995	5
`set-04.zip`	4,987	13
`set-05.zip`	4,996	4
`set-06.zip`	4,994	6
`set-07.zip`	4,996	4
Total	39,954	46

All 40,000 videos have a validated terminal annotation file. A terminal failure record describes an unsuccessful annotation and is not a usable training annotation. The completion counts include successful fallback runs. No dataset viewer or tabular preview is configured.

Download and read annotations

Download only the splits you need. For example:

hf download YukkiLancer/OV4sKSFc archives/set-05.zip \
  --repo-type dataset --local-dir ./OV4sKSFc
hf download YukkiLancer/OV4sKSFc SHA256SUMS \
  --repo-type dataset --local-dir ./OV4sKSFc
cd OV4sKSFc
sha256sum -c SHA256SUMS --ignore-missing
unzip archives/set-05.zip -d set-05

The ZIPs use ZIP64 and store the already gzipped JSON files without another compression pass. SHA256SUMS verifies each complete ZIP. To inspect a file without extracting the ZIP:

import gzip
import json
from zipfile import ZipFile

with ZipFile("archives/set-05.zip") as archive:
    name = next(name for name in archive.namelist() if name.endswith(".json.gz"))
    with archive.open(name) as compressed:
        annotation = json.load(gzip.GzipFile(fileobj=compressed))
print(annotation["status"], annotation["video"]["video_id"])

source-index/set-XX.jsonl maps each annotation ID to its source video name, caption, pinned upstream ZIP member, raw-video SHA-256, annotation SHA-256, and terminal status. schemas/annotation.schema.json describes the canonical annotation contract. The source index is download metadata; it is not a Dataset Viewer table.

Download corresponding raw videos

Download the script and source index, then choose one or more split IDs:

from huggingface_hub import snapshot_download

snapshot_download(
    "YukkiLancer/OV4sKSFc",
    repo_type="dataset",
    local_dir="OV4sKSFc",
    allow_patterns=["download_raw_videos.py", "requirements-download.txt",
                    "source-index/*.jsonl"],
)

cd OV4sKSFc
python -m pip install -r requirements-download.txt
python download_raw_videos.py --splits 5 6 7 --output-dir ./raw-videos --workers 8

Omit --splits to download all eight sets, or use --limit-per-split 1 for a small check. The complete selection totals about 342 GB of raw MP4 bytes. Downloads go to raw-videos/set-XX/. The script reads only the selected ZIP byte ranges from the pinned upstream revision, verifies ZIP CRC32 and each video's SHA-256, and safely resumes after interruption. Rerun the same command to verify and skip existing videos. Access to the upstream dataset and enough local disk space are required.

Provenance and use

The selection consists of eight disjoint sets of 5,000 videos. release.json records the selection digest and the source annotation generation for each set. The annotations use a 6 FPS sampling profile with Qwen, SAM, and VidEoMT stages. They can contain model errors, missed objects, and identity mistakes. The original source captions and videos are attributed to OpenVid-1M, whose page describes its license and the additional terms of the underlying video sources. No separate license for these generated annotations is declared here.

Downloads last month: 32

Total file size:

59.8 GB