flashdata-dev's picture
Add typed Parquet index for the dataset viewer
17d6784 verified
|
Raw History Blame Contribute Delete
6.59 kB
metadata
pretty_name: VideoUFO Everyday Activities — FlashData Sample
license: cc-by-4.0
language:
  - en
size_categories:
  - n<1K
tags:
  - video
  - video-text
  - video-captioning
  - youtube
  - custom-video-dataset
  - flashdata
  - videoufo
configs:
  - config_name: default
    data_files:
      - split: sample
        path: viewer/sample.parquet

VideoUFO Everyday Activities — FlashData Sample

Food preparation, personal care, music and everyday actions for testing video understanding and multimodal data pipelines.

12 downloadable MP4 clips · 8.55 MB of video · English captions · CC BY 4.0

Request a custom video dataset · Try the YouTube Download API

Preview frames from the twelve sample clips

What you can download

This small selection from VideoUFO contains H.264 video files, brief and detailed source captions, source timestamps, six upstream quality scores, measured file properties and SHA-256 checksums. FlashData selected and packaged these samples; the videos and source annotations were published by the VideoUFO authors. FlashData transcoded the upstream MPEG-4 Part 2 video to H.264 for browser playback, without trimming, resizing or changing the frame rate. Existing AAC audio was copied.

Property Value
Clips / distinct source video IDs 12 / 12
Total clip duration 88.79 seconds
Clip duration range 3.17–23.23 seconds
Split sample
Media H.264 MP4 with existing AAC audio and overlays
Captions English source descriptions, not audio transcripts
Checks Every selected file checksum verified and fully decoded

Load the sample

Install the Hugging Face datasets library:

pip install datasets
from datasets import Video, load_dataset

sample = load_dataset("flashdata-dev/videoufo-everyday-activities-sample", split="sample")
# Keep file references without requiring a video decoder for this example.
sample = sample.cast_column("video", Video(decode=False))
print(sample[0]["caption"])
print(sample[0]["video"])

You can also download individual MP4s from the Files and versions tab. The data/metadata.jsonl file links each record to its adjacent MP4. The viewer/sample.parquet index contains the same metadata and typed video references pinned to an immutable media commit.

Fields

Field Meaning
video / file_name Typed video reference / raw metadata's relative MP4 path
clip_id, source_video_id, source_video_url Preserved VideoUFO ID and corresponding YouTube source reference
topic Original VideoUFO topic, not a verified action label
caption, detailed_caption Brief and detailed annotations copied from VideoUFO
source_start_time, source_end_time Original source-video interval; these are not local clip offsets
source_start_seconds, source_end_seconds Numeric forms of the original timestamps
duration_seconds, width, height, has_audio Measured properties of the packaged file
byte_size, sha256 Integrity information for the published H.264 file
source_byte_size, source_sha256 Integrity information for the original upstream MP4
video_codec h264 for browser playback
aesthetic_quality, background_consistency, dynamic_degree, imaging_quality, motion_smoothness, subject_consistency Unmodified upstream VBench scores
source_dataset, source_revision, license Source and license provenance

Selection and limitations

These 12 clips were manually selected from the first 256 MP4 members of upstream archive 1, with frames inspected at 10%, 50% and 90% of each selected clip. This is a convenience sample for inspecting a delivery format and testing loaders. It is not a representative benchmark or a production training set.

The upstream videos are compressed and may contain embedded text, watermarks or other overlays. Captions and quality scores are inherited, may contain errors, and have not been relabeled as ground truth. No train/test split is provided. The source reference does not establish that a video remains available on YouTube.

Need a larger or different collection?

FlashData offers custom video dataset collection and a YouTube Download API for delivering media to your cloud storage. Bring video IDs, channel IDs or a dataset brief. Describe the topics, languages, duration, resolution, volume and whether you need metadata or available captions.

For a HowTo100M-style source-ID list, ask about collecting available source videos and recording delivery outcomes in your bucket. This sample demonstrates a package format; it is not a HowTo100M mirror or a FlashData API download run.

Attribution and license

Source: VideoUFO: A Million-Scale User-Focused Dataset for Text-to-Video Generation, by the VideoUFO authors. The upstream repository publishes its videos and dataset under CC BY 4.0. This package retains that license.

Upstream revision: 9ec8b5d8c2dbc14e2071c6d5e7cbbd16c5b532c6. FlashData's changes are selection, field normalization, measured media properties, integrity checks, H.264 transcoding and preview images. Source captions are unchanged; video bytes changed during encoding. See ATTRIBUTION.md, provenance.json and LICENSE. No endorsement by the original authors is implied.