lioryariv/vchitect-100k

vchitect video subset (100k clips)

A 100,000-clip subset of the Vchitect T2V DataVerse video collection, packed as tar shards for transfer.

Contents

Path	Description
videos/shard-00000.tar … shard-00168.tar	169 tar shards, ~2 GiB each, 343.3 GiB total. Each contains ~600 .mp4 files stored flat by their original basename (e.g. 0000101587.mp4).
manifests/resolved_selected_videos.txt	Original 100k-line manifest of selected clips (absolute paths at the staging location).
manifests/resolved_selected_videos.source.txt	The same list against the original source location.
shard_index.json	JSON array of 169 lists; element i holds the filenames packed into shard-{i:05d}.tar. Use it to find which shard a given clip lives in.

The manifests contain paths only — no captions or other annotations are included.

Download

Everything (343 GiB):

hf download lioryariv/vchitect-100k --repo-type dataset --local-dir ./vchitect

A few shards only:

hf download lioryariv/vchitect-100k --repo-type dataset \
  --include "videos/shard-0000*.tar" --local-dir ./vchitect

Unpack:

mkdir -p videos && for f in vchitect/videos/*.tar; do tar -xf "$f" -C videos; done

Streaming without unpacking

The shards are plain tars of .mp4 files, so webdataset reads them directly:

import webdataset as wds

url = "https://huggingface.co/datasets/lioryariv/vchitect-100k/resolve/main/videos/shard-{00000..00168}.tar"
ds = wds.WebDataset(url).to_tuple("mp4")

For a private repo, pass your token — e.g. export HF_TOKEN and use hf_hub_download, or add an Authorization: Bearer $HF_TOKEN header.

Provenance and licensing

These clips originate from the Vchitect T2V DataVerse dataset; the underlying videos are subject to their original sources' terms. This repo is a private working copy for research collaboration and is not a redistribution of the upstream dataset. Check the upstream license before sharing further or using the data in a published artifact.