docs/video_sources.md

Video sources and local reconstruction

This repository publishes annotations only. Download upstream videos into a local directory after reviewing the upstream terms and access requirements.

Component	Upstream repository	Pinned revision inspected for this release
QA annotations	CASIA-IVA-Lab/ThinkStream	94a9d82bc74864ba6b705b52ffa3a2d4bc9733e8
LLaVA video packages	lmms-lab/LLaVA-Video-178K	6d8c562dc26d70042a0d9704d1cae58c94b89098
Kinetics-700 video package	omni-research/Tarsier2-Recap-585K	5cffa8342bbd9900fb50fdb7fd1b5f2cd1778ac4

The last source is access-gated. Request access on its Hugging Face page and log in locally before downloading. Keep access tokens out of scripts and command history. Pinned repository revisions document upstream file listings; they are not a claim that every local video was byte-hashed against that revision.

Required LLaVA subsets

The release references these five subsets, with their original names:

0_30_s_youtube_v0_1
0_30_s_academic_v0_1
1_2_m_academic_v0_1
1_2_m_activitynetqa
2_3_m_academic_v0_1

metadata/video_manifest.parquet is the exact list of required files. Archives may contain additional videos. Avoid downloading the entire upstream repository if only these annotations are needed.

Example, sequential download of the required archive groups:

from huggingface_hub import snapshot_download

subsets = [
    "0_30_s_youtube_v0_1", "0_30_s_academic_v0_1",
    "1_2_m_academic_v0_1", "1_2_m_activitynetqa", "2_3_m_academic_v0_1",
]
snapshot_download(
    repo_id="lmms-lab/LLaVA-Video-178K", repo_type="dataset",
    revision="6d8c562dc26d70042a0d9704d1cae58c94b89098",
    allow_patterns=[f"{subset}/*.tar.gz" for subset in subsets],
    local_dir="downloads/LLaVA-Video-178K", max_workers=1,
)

Inspect each archive's member paths, then extract each archive under its matching subset directory. For example, for one archive:

mkdir -p /path/to/videos/LLaVA-Video-178K/0_30_s_youtube_v0_1
tar -xzf downloads/LLaVA-Video-178K/0_30_s_youtube_v0_1/0_30_s_youtube_v0_1_videos_1.tar.gz \
  -C /path/to/videos/LLaVA-Video-178K/0_30_s_youtube_v0_1

Historical annotations contain .../<subset>/data/<source-folder>/.... The archive may instead extract .../<subset>/<source-folder>/.... The preparation script tries both forms, so a machine-specific symlink is unnecessary. liwei_youtube_videos, academic_source, and ActivityNet-QA are inherited upstream directory names.

Kinetics-700 package

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="omni-research/Tarsier2-Recap-585K", repo_type="dataset",
    revision="5cffa8342bbd9900fb50fdb7fd1b5f2cd1778ac4",
    allow_patterns=["Kinetics-700/videos.tar.part-*", "Kinetics-700/metadata.json"],
    local_dir="downloads/Tarsier2", max_workers=1,
)

The inspected package has 13 numbered videos.tar.part-* files. After obtaining all parts, concatenate in filename order and extract the tar stream:

mkdir -p /path/to/videos/Kinetics-700
cat downloads/Tarsier2/Kinetics-700/videos.tar.part-* \
  | tar -xf - -C /path/to/videos/Kinetics-700

The archive contains videos/<clip-id>.mp4. Clip filenames such as <video-id>_000179_000189.mp4 identify an upstream clip; the annotation interval is relative to that clip, not the original full YouTube timeline. Re-downloading from another source can change clip boundaries or availability, so verify correspondence.

Expected layout

videos/
├── LLaVA-Video-178K/
│   ├── 0_30_s_youtube_v0_1/liwei_youtube_videos/...
│   ├── 0_30_s_academic_v0_1/academic_source/...
│   ├── 1_2_m_academic_v0_1/academic_source/...
│   ├── 1_2_m_activitynetqa/ActivityNet-QA/...
│   └── 2_3_m_academic_v0_1/academic_source/...
└── Kinetics-700/videos/...

Validate path mapping without writing output:

python scripts/prepare_local_data.py --video-root /path/to/videos --config current --check-only

Use --llava-root and --kinetics-root if sources are stored separately. The check does not download, decode or modify videos. File existence alone cannot guarantee correct clip identity, decodability, or the validity of every time interval.