license: other
pretty_name: MCWM VPT 7.x Canonical Dataset
tags:
- minecraft
- reinforcement-learning
- video
- world-model
MCWM VPT 7.x Canonical Dataset
This dataset is a reproducible 5,460-episode subset of the OpenAI Video PreTraining (VPT) 7.x contractor demonstrations, converted to the canonical MCWM action and timestamp schema.
Contents
- 5,460 MP4 videos at 640x360
- 30,133,637 frames and 28,202,719 aligned canonical action ticks
- 418.45 hours of video
- contractor-grouped splits: 4,767 train, 296 validation, 397 test
- exact frame PTS, per-episode manifests, checksums, and an audit report
The 1,011 selected source paths that returned HTTP 404 are recorded
under selection/skipped-missing-relpaths.txt. Another 29 source
action files that failed UTF-8/JSON/timestamp validation are recorded under
selection/skipped-invalid-actions-relpaths.txt. The published canonical audit
has no split leakage and no unresolved issues.
Layout and use
Pass the downloaded canonical/ directory as the MCWM data root. Manifest
video paths are relative and resolve to videos/<recorder-version>/<file>.mp4.
Source and attribution
Derived from the OpenAI VPT contractor data: https://github.com/openai/Video-Pre-Training
OpenAI describes the contractor data as open sourced for research use. The upstream repository also states that Minecraft is Microsoft's intellectual property and that OpenAI does not purport to license Minecraft intellectual property. Users are responsible for reviewing the upstream terms and ensuring their intended use is permitted. This processed dataset does not grant rights beyond those applicable to the upstream material.