license: mit
task_categories:
- reinforcement-learning
pretty_name: pan-2 VPT (64px episodes + packed shards)
size_categories:
- 100GB<n<1TB
tags:
- minecraft
- vpt
- openai
- numpy
pan-2 VPT
Cleaned OpenAI VPT contractor demos, packed for Infatoshi/pan-2.
This is not a re-dump of the original VPT Azure blobs, and it is not the same layout as zhwang4ai/OpenAI-Minecraft-Contractor (jsonl-only), p-doom/openai-minecraft-dataset (Grain/ArrayRecord tars), or TESS-Computer/minecraft-vla-stage1 (parquet JPEGs at 640x360 / 5 Hz).
| Episodes | 1625 validated stems (27 dropped) |
| Frames | 8,155,382 (~113.3 h at 20 Hz) |
| Image | uint8 [T, 64, 64, 3] |
| Actions | float32 [T, 25] (23 buttons + camera dx/dy in [-1, 1]) |
| Raw video | H.264 640x360 @ 20 fps, sibling jsonl |
Layout
raw/<stem>.mp4
raw/<stem>.jsonl # original VPT per-tick action dicts
episodes/<stem>.img.npy # (T, 64, 64, 3) uint8
episodes/<stem>.act.npy # (T, 25) float32
shards/manifest.jsonl
shards/shard-XXXXX.frames.npy
shards/shard-XXXXX.act.npy
meta/README.md
meta/cleanup_summary.json
meta/episodes_manifest.jsonl
Episodes never straddle shards. manifest.jsonl starts with a header:
{"type":"header","version":1,"image_size":64,"act_dim":25,"n_shards":24,"total_frames":8155382,"total_episodes":1625}
then one segment row per episode (shard, stem, offset, n_frames, has_act).
Load shards
import json
from pathlib import Path
import numpy as np
root = Path("shards")
header = json.loads(root.joinpath("manifest.jsonl").read_text().splitlines()[0])
frames = np.load(root / "shard-00000.frames.npy", mmap_mode="r") # (N, 64, 64, 3) uint8
acts = np.load(root / "shard-00000.act.npy", mmap_mode="r") # (N, 25) float32
Rebuild shards from episodes with the pan-2 repo:
uv run python scripts/build_shards.py --source episodes --episodes-dir episodes --out shards
Source
OpenAI Video PreTraining contractor data (Baker et al., 2022). Original index/blobs: https://openaipublic.blob.core.windows.net/. Some origin mp4s 404; this tree is the surviving cleaned 1625-stem subset used for pan-2.
Action columns: see src/pan2/actions.py in the code repo.
License
MIT, same as the upstream VPT contractor release. Cite OpenAI VPT if you use this.