MineJEPA Treechop 4.6M
MineJEPA Treechop 4.6M is an action-conditioned latent world model trained on
MineRL Treechop human demonstrations. It predicts future latent states
from a current RGB observation and a sequence of MineRL action vectors. It is
not a pixel/video generator and is not a Minecraft-playing agent.
Model summary
- Architecture: convolutional RGB encoder + residual action-conditioned MLP
dynamics + deterministic EMA target encoder.
- Parameters: approximately 4.6M trainable parameters.
- Input: 64x64 RGB MineRL POV frames and 10-dimensional flattened MineRL
action vectors.
- Training: 200k updates, batch size 32, horizons 1/3/5/10, with an
action-contrastive objective at horizons 3 and above.
- Recommended artifact:
checkpoints/checkpoint-200k.safetensors.
Training data and split
The model was trained on 209 MineRLTreechop-v0 human demonstration
trajectories. The converted local Lance dataset contained 446,574 16-frame
windows. An 80/20 whole-episode split, seed 17, reserved 42 trajectories
from training for evaluation.
This repository does not redistribute MineRL data, converted Lance shards, raw
frames, or action trajectories. Users must download source data through MineRL
and comply with its terms before reproducing conversion or training.
The matching
MineJEPA Treechop conversion manifest
records the exact conversion schema, action encoding, split, and local
reproduction command used for this release.
Held-out results
The following results compare the recorded action sequence with mismatched
action sequences while holding the start observation and true future target
fixed. Metrics aggregate three deterministic samples of 512 windows from the
same 42 held-out episodes.
| H=10 checkpoint | Action margin | Action win | Future-latent R@1 | R@5 |
|---|
| 140k | 0.0955 ± 0.0011 | 89.2% | 63.4% | 85.1% |
| 200k | 0.1209 ± 0.0030 | 91.2% | 64.1% | 84.1% |
action margin is recorded-target cosine minus mismatched-action-target
cosine. It measures whether the model's predicted future is more compatible
with the action sequence that actually occurred.
Intended use
Use this checkpoint for research into latent dynamics, counterfactual action
diagnostics, feature retrieval, and offline planning experiments on compatible
MineRL Treechop observations and action encoding.
Limitations
- The model has no pixel decoder and cannot generate Minecraft video.
- It was evaluated on one task and one episode-level split only.
- It is not validated for real-time environment control.
- Its H=10 goal-action retrieval is only 3.1% R@1 among 256 candidate action
chunks. It is therefore not a reliable large-library goal planner.
- Do not use it outside research/educational contexts without independent
validation.
Loading
Install the repository at the release commit, download config.json and
checkpoint-200k.safetensors, then load the full state dict into
MineJEPAWorldModel(action_dim=10, latent_dim=512, hidden_dim=1024) using
safetensors.torch.load_file.
For exact commands and fine-tuning guidance, see the GitHub repository's
docs/MINEJEPA_HF_RELEASE.md.
Citation
If you use this release, cite the MineJEPA-SWM repository and link to its
evaluation reports. This is an independent research implementation and is not
an official MineRL or Stable World Model checkpoint release.
