---
license: other
pretty_name: GameVerse
language:
- en
task_categories:
- text-to-video
- image-to-video
- video-to-video
size_categories:
- 10K<n<100K
tags:
- text
- video
- tabular
- csv
- gameplay
- game-video
- camera-pose
- world-modeling
configs:
- config_name: gameverse_dataset
  data_files:
  - split: train
    path: gameverse_dataset.csv
---

# GameVerse

This repository releases reproducible URL/timestamp metadata and paired
camera-pose/caption annotations for GameVerse gameplay clips.

The MP4 clips are not redistributed. Each row in `gameverse_dataset.csv`
provides the source YouTube URL, fps from the source JSONL `video_fps` field,
frame range, timestamp range, game name, style keywords, style sentence, and
logical annotation paths.

## Files

- `gameverse_dataset.csv`: 69,678 gameplay clips.
- `annotations/pose.tar`: all pose files.
- `annotations/caption.tar`: all caption files.
- `metadata/annotation_archives.csv`: size and SHA-256 checksums for the two
  annotation archives.

## Annotation paths

The CSV uses logical paths:

- `pose_path`: `annotation/pose/<clip_id_without_mp4>.npz`
- `caption_path`: `annotation/caption/<clip_id_without_mp4>.json`

These paths are the member names inside `annotations/pose.tar` and
`annotations/caption.tar`. For example, if `pose_path` is
`annotation/pose/-57bKMZQh18_0180010_0187210.npz`, that file is stored as the
same member path inside `annotations/pose.tar`.

Caption copies in this release remove `annotation_meta.model` and
`annotation_meta.machine_id`; semantic caption fields are kept.

## Paper

[GameVerse: A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling](https://arxiv.org/abs/2610.37257) (NeurIPS 2026).