Blame for README.md

---
license: other
license_name: mixed-third-party-licenses
license_link: LICENSE
configs:
  - config_name: events
    default: true
    data_files:
      - split: events
        path: metadata/events.parquet
      - split: sample_events
        path: samples/metadata/events.parquet
  - config_name: videos
    data_files:
      - split: videos
        path: metadata/videos.parquet
      - split: sample_videos
        path: samples/metadata/videos.parquet
---

# MEV: A Multi-Event Video Dataset for Long-Take Generation

<p align="center">
  <a href="https://openreview.net/forum?id=blH6L8WmTv"><img src="https://img.shields.io/badge/Paper-OpenReview-B31B1B?style=flat&amp;labelColor=344054" height="20" alt="Paper · OpenReview"></a>&nbsp;
  <a href="https://github.com/SoftPointer/MEV"><img src="https://img.shields.io/badge/GitHub-Code-181717?style=flat&amp;labelColor=344054&amp;logo=github&amp;logoColor=white" height="20" alt="GitHub · Code"></a>&nbsp;
  <a href="https://mev-dataset.github.io/"><img src="https://img.shields.io/badge/Project-Page-2563EB?style=flat&amp;labelColor=344054" height="20" alt="Project Page"></a>&nbsp;
  <a href="#citation"><img src="https://img.shields.io/badge/Citation-BibTeX-BE185D?style=flat&amp;labelColor=344054" height="20" alt="Citation · BibTeX"></a>&nbsp;
  <a href="#citation"><img src="https://img.shields.io/badge/NeurIPS-2026-7C3AED?style=flat&amp;labelColor=344054" height="20" alt="NeurIPS 2026"></a>
</p>

MEV is a dataset for long-take multi-event video generation, with 8k real-world
videos and 25k temporally grounded events. It provides Global and Event
annotations, including multi-granularity captions, camera descriptions, semantic
attributes, and six VBench quality scores.

The canonical nested annotation file is `annotations/mev.json`. The primary
machine-readable table is `metadata/events.parquet`, where each row corresponds
to one event segment.

## Repository Layout

- `annotations/mev.json`: canonical full annotation file.
- `metadata/events.parquet`: full event-level table.
- `metadata/events.csv` and `mev.csv`: CSV mirrors for inspection.
- `metadata/videos.parquet`: UUID-level index.
- `raw/720p/raw-*.tar`: 8,016 original videos, stored as `<uuid>.mp4`.
- `raw/1080p/raw-*.tar`: 1,896 optional higher-resolution originals.
- `videos/mev-*.tar`: full video segment tar shards. Tar members preserve the original mp4 filenames.
- `samples/`: reviewer-friendly sample subset with matching annotation, metadata, and tar shard(s).

## Loading Metadata

```python
from datasets import load_dataset

events = load_dataset("MultiEventVideo/MEV", split="events")
videos = load_dataset("MultiEventVideo/MEV", "videos", split="videos")
sample_events = load_dataset("MultiEventVideo/MEV", split="sample_events")
```

The default `events` configuration contains the `events` and `sample_events`
splits. The `videos` configuration contains `videos` and `sample_videos`.
These are table and inspection-subset names, not experimental train/test splits.

Each event row includes `shard_path`, `tar_member`, and `media_ref`. For example:

```text
videos/mev-000000.tar::<original-segment-filename>.mp4
```

`has_cooking` is a boolean indicating food or beverage preparation, or active
cooking/heating, during at least part of the event.

## Download and Extract Videos

Loading the metadata with `load_dataset` does not download the video archives.
The following commands use a macOS or Linux shell.

### Download the dataset

Install the [Hugging Face CLI](https://huggingface.co/docs/huggingface_hub/guides/cli)
and download the repository to a local `MEV` directory:

```bash
python -m pip install -U huggingface_hub
hf download MultiEventVideo/MEV --repo-type dataset --local-dir ./MEV
cd MEV
```

The complete download is approximately 227 GB. Run the commands below from this
`MEV` directory. Individual archives can also be downloaded from the repository's
Files tab, preserving the same directory layout.

### Extract the original and event videos

The `.tar` archives contain MP4 files and have no additional gzip compression.
Create an `extracted/` subdirectory within each archive directory and extract the
videos there. The tar files remain in their existing locations:

```bash
for directory in raw/720p raw/1080p videos; do
  mkdir -p "$directory/extracted"
  for archive in "$directory"/*.tar; do
    tar -xf "$archive" -C "$directory/extracted"
  done
done
```

After running these commands, the directories contain:

```text
raw/720p/
  raw-*.tar
  extracted/<uuid>.mp4                  # 8,016 primary original videos
raw/1080p/
  raw-*.tar
  extracted/<uuid>.mp4                  # 1,896 available alternate originals
videos/
  mev-*.tar
  extracted/<original-event-filename>.mp4 # 25,262 event videos
```

Here, `raw/` holds original videos, while `videos/` holds Event videos.

Open the extracted MP4 files in a video player. Original filenames are preserved:
match raw videos by `uuid` and event videos by each annotation's `filename`.
The 1080p directory contains alternate originals, not additional dataset UUIDs.
Keeping the archives and all extracted full-data videos uses approximately 450 GB.

To inspect a single archive before extracting it:

```bash
tar -tf videos/mev-000000.tar
```

Videos can also be read directly from their tar archives using the Python example
under **Original video access**, without extracting the complete archives.

## Sample Subset

The `samples/` directory contains a smaller subset intended for reviewer
inspection. It is drawn from the full MEV release at the UUID level; all events
from selected UUIDs are preserved.

Selection includes all videos with saved human-review records from the web
annotation interface that are present in this release: 226 original
videos and 721 events. Annotations are copied from the current full
dataset. Selection is at the video level; it does not imply that every field was
individually reviewed. `samples/uuids.txt` records the exact membership.
`samples/raw/720p/` contains the original videos, and `samples/videos/` contains
their event videos. This is an inspection subset, not a random population sample
or the experimental train/test split.

## License

This repository uses `mixed-third-party-licenses`.

Original Mixkit, YouTube, and other third-party videos remain subject to their respective platform terms and rights-holder licenses. The dataset authors do not claim ownership of third-party video assets and do not re-license them under CC BY, MIT, Apache, or any other open-source license.

The official released video package is directly downloadable for research,
evaluation, and model training, subject to applicable source-specific terms
described in `LICENSE`. Optional restricted or premium source variants are not
automatically part of the public package. Preserve available source attribution
and provenance. The repository's existing `LICENSE` is authoritative.

Annotations, labels, splits, benchmark metadata, and other author-created materials are released under the Creative Commons Attribution 4.0 International License, CC BY 4.0.

See the `LICENSE` file for full details.

## Original video access

The dataset contains 8,016 original videos and 25,262 events.

The Global table (`metadata/videos.parquet`) includes `raw_720p_shard_path`,
`raw_720p_tar_member`, `raw_720p_media_ref`, `raw_720p_bytes`, and `raw_720p_sha256`.
The full table also links available alternate originals with the corresponding
`raw_1080p_*` columns. Empty alternative references indicate that the table does
not provide that variant. The sample supplies the primary original videos.
Resolution directory names preserve the source storage groups; actual media
dimensions can vary. Every tar preserves the original MP4 bytes and filenames.

All archives are uncompressed tar shards, each at most 5 GiB. To read an
original video and its first event directly from the downloaded archives,
install the table-reading dependencies:

```bash
python -m pip install pandas pyarrow
```

Then run this Python example from the downloaded `MEV` directory:

```python
from pathlib import Path
import tarfile
import pandas as pd

root = Path(".")
videos = pd.read_parquet(root / "metadata/videos.parquet")
events = pd.read_parquet(root / "metadata/events.parquet")
video = videos.iloc[0]
event = events[events.uuid == video.uuid].iloc[0]

def read_media(reference):
    archive, member = reference.split("::", 1)
    with tarfile.open(root / archive) as tar:
        return tar.extractfile(member).read()

raw_bytes = read_media(video.raw_720p_media_ref)
event_bytes = read_media(event.media_ref)
```

Join the Global and Event tables on `uuid`. Each Event row stores its original
video timestamps, an event ID starting at 1, its tar reference, byte size, and
SHA-256. The HF table split names identify data views, not experimental train/test
assignments.

## Citation

If you use MEV in your research, please cite:

```bibtex
@misc{mev2026,
  title  = {MEV: A Multi-Event Video Dataset for Long-Take Generation},
  author = {Zhu, Peiyuan and Xie, Shaoan and Shen, Yifan and Tian, Wen and
            Li, Zijian and Chen, Guangyi and Zhang, Kun},
  year   = {2026},
  url    = {https://mev-dataset.github.io/}
}
```