Paper and Code

Associated with the paper: WikiVideo (https://arxiv.org/abs/2504.00939)

Associated with the github repository (https://github.com/alexmartin1722/wikivideo)

Download instructions

The dataset can be found on huggingface. However, you can't use the datasets library to access the videos because everything is tarred. Instead you need to locally download the dataset and then untar the videos (and audios if you use those).

Step 1: Install git-lfs

The first thing you need to do is make sure that git-lfs is installed, otherwise you won't be able to pull the video and audio tar files.

git lfs install

Step 2: Clone the dataset

After enabling git-lfs, you can now pull the dataset from huggingface.

git clone https://huggingface.co/datasets/hltcoe/wikivideo

I would also tmux this because it might take a while.

Step 3: Untar the videos

Videos from MultiVENT2.0 (WikiVideo 2024)

These videos range from 2015-2024.

# untar the videos 
tar -xvzf videos.tar.gz -C data/videos
# untar the audios
tar -xvzf audios.tar.gz -C data/audios

Videos from the year 2025 (WikiVideo25)

These videos are for the year 2025. They are also the videos used in the MAGMaR shared task at ACL 2026.

# untar the videos 
tar -xvzf videos_2025.tar.gz -C data/videos
# untar the audios
tar -xvzf audios_2025.tar.gz -C data/audios

Finish

Now you should be done. You will see a annotations folder in the huggingface repo, but this also exists in the data/ folder already in the data/wikivideo directory.

Dataset Format

In the data/wikivideo directory, you will find the file final_data.json which has the articles. This file is formatted as a dictonary:

{
    "Wikipedia Title": {
        "claims": [["claim1", "claim2", ...], ...], 
        "original_article": ["sent1", "sent2", ...],
        "claims_to_supporting_videos": {
            "claim1": {
                "supporting_videos": ["video_id1", "video_id2", ...],
                "videos_modalities": {
                    "video_id1": {
                        "video": true,
                        "audio": false,
                        "ocr": true
                    },
                    ...
                }
            }
        },
        "article": "The article text",
        "query_id": "query id",
        "videos": {
            "video_id1": {
                "anon_scale_id": "XXX",
                "language": "english",
                "video_type": "Professional | Edited | Diet Raw | Raw",
                "relevance": 3
        }
    },
    ...
}

In this json, you see that the top level key is the Wikipeda Article Title. Each other key is defined as follows:

• claims: The claims from the original article
• original_article: The original article split into sentences
• claims_to_supporting_videos: A mapping from each claim to the videos that support it along with the modalities present in each video
• article: This is the human written article on the topic using the video data.
• query_id: This is the query id for the article from the MultiVENT 2.0 dataset. This will be helpful when doing RAG experiments.
• videos: Metadata about each video used in the article

RAG Data

The videos that we used as the distractor set (and also include these videos in videos.tar.gz) can be found here MultiVENT 2.0 (https://huggingface.co/datasets/hltcoe/MultiVENT2.0)

The relevance judgments for this data can be found in annotations/qrels.trec

---

license: apache-2.0