Hugging Face's logo

Datasets:
JFJFJFen
/
mv-raw-subset

Modalities:
Video
Tags:
video
video-retrieval
multilingual
License:
Dataset card Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset contains news video clips collected from YouTube (MultiVENT-raw). Access is granted for non-commercial research use only. By requesting access you agree not to redistribute the videos and to use them solely for research.

Log in or Sign Up to review the conditions and access this dataset content.

mv-raw-subset

Evaluation-pool subset of MultiVENT-raw: the 17,728 video chunks that enter the top-100 reranking pools of two dense first-stage retrievers over the 222 MultiVENT-raw evaluation queries.

Contents

  • videos/shard_NNNNNN.tar (662 tars, ~607 GB total): flat members <chunk_id>.mp4 (H.264/AAC) + <chunk_id>.json (per-chunk metadata). Shard numbering matches the original MultiVENT-raw release; each tar here contains only the pool members of that source shard.
  • catalog_pool.csv (17,728 rows): chunk_id, video_id, chunk_index, chunk_count, shard_index, duration_sec, chunk_start_sec, chunk_end_sec, size_bytes, vcodec, acodec. shard_index N maps to videos/shard_%06d.tar.
  • mvraw_eval_pool_omniemb01mv_video_ids.txt (12,792): top-100 union, OmniEmbed-v0.1-multivent first stage.
  • mvraw_eval_pool_omniemb01_video_ids.txt (10,544): top-100 union, OmniEmbed-v0.1 (base) first stage.
  • mvraw_eval_pool_both_union_video_ids.txt (17,728): union of the two (overlap 5,608); this is the id set shipped in videos/.

Pool construction

222 evaluation queries (MultiVENT-raw annotations release), dense retrieval over all 143,288 chunks, top-100 per query, union over queries. Chunks are ~5 min max segments of source videos (videos over 300 s were split; 99% of sources are single-chunk). Mean chunk duration in this pool: 143 s.

Loading example 

import tarfile, io, json
from huggingface_hub import hf_hub_download

path = hf_hub_download("JFJFJFen/mv-raw-subset", "videos/shard_000000.tar", repo_type="dataset")
with tarfile.open(path) as tf:
    for m in tf:
        if m.name.endswith(".json"):
            meta = json.load(tf.extractfile(m))
        elif m.name.endswith(".mp4"):
            mp4_bytes = tf.extractfile(m).read()

Provenance and license

Chunks are repacked verbatim (no re-encoding) from the MultiVENT-raw release, which was collected from public YouTube uploads with yt-dlp. Rights to the underlying videos remain with the original uploaders; this subset is redistributed for non-commercial research only. If you are a rights holder and want content removed, open a discussion on this repo.
Downloads last month
4
Total file size:
607 GB