asdfo123/ForgeWM-data

ForgeWM Training Data

Pre-encoded latents for training ForgeWM.

This is a re-packaging of the GameFactory GF-Minecraft dataset, encoded into Wan2.1 VAE latents and sharded as LMDB for direct use by ForgeWM training scripts. The underlying gameplay videos are from GameFactory — we just made them training-ready.

Quick Start

huggingface-cli download asdfo123/ForgeWM-data \
  --local-dir ./data/action_lmdb --repo-type dataset

Then point ForgeWM configs to ./data/action_lmdb.

What's Inside

Field	Value
Clips	4000
Shards	10 LMDB files
Latent shape	(21, 16, 44, 80) — Wan2.1 VAE, 84 frames @ 352×640 decoded
Keyboard	one-hot W/S/A/D
Mouse	[yaw, pitch], normalized
Total size	~89 GB

Processing note: GameFactory's pitch convention (+pitch = look-down) is flipped to MG2's (mouse[0] > 0 = look-up) at encoding time.

Citation

@misc{forgewm2026,
  title={ForgeWM: A Reproducible Training Recipe for Action-Controllable World Models},
  author=ForgeWM Team,
  year={2026},
  url={https://github.com/asdfo123/ForgeWM}
}

Please also cite GameFactory (the underlying data):

@misc{yu2024gamefactory,
      title={GameFactory: Creating New Games with Generative Interactive Videos}, 
      author={Yu, Jiwen and Qin, Yiran and Wang, Xintao and Wan, Pengfei and Zhang, Di and Liu, Xihui},
      year={2025},
      eprint={2501.08325},
      archivePrefix={arXiv},
}

License

Apache 2.0 for this repackaging. The underlying GF-Minecraft data follows GameFactory's license — please consult their terms before redistribution. ```