---
license: mit
tags:
  - reinforcement-learning
  - embodied-ai
  - behavior-cloning
  - crafter
  - craftax
  - video-language
---

# CrafterDojo v1 weights

Pretrained models from **CrafterDojo: An Integrated Research Suite for Building Open-Ended Embodied Agents in Crafter**
([paper](https://arxiv.org/abs/2508.13530), [code](https://github.com/frechele/CrafterDojo), [project](https://sites.google.com/view/crafterdojo)).

| File | Model | Notes |
|---|---|---|
| `vpt/tiny.safetensors` | CrafterVPT-tiny | behavior cloning, last checkpoint |
| `vpt/base.safetensors` | CrafterVPT-base | behavior cloning, last checkpoint |
| `vpt/large.safetensors` | CrafterVPT-large | behavior cloning, last checkpoint |
| `crafterclip/crafterclip.safetensors` | CrafterCLIP | video–caption encoder (MineCLIP architecture, ViT-B/16 backbone) |
| `steve1/steve1.safetensors` | CrafterSteve-1 | instruction-following policy (LoRA on CrafterVPT-base) |
| `steve1/steve1_prior.safetensors` | CrafterSteve-1 prior | text → visual-goal CVAE prior |
| `vpt/*.yaml` | architecture configs | same files as `models/vpt/*.yaml` in the code repository |

All models are trained on the `sv_gh_pp-10B_20ep` CrafterPlay mixture (Survival + Gatherer experts) in Craftax-Classic.
Each `.safetensors` file records its training run in the header metadata.

## Usage

```bash
git clone https://github.com/frechele/CrafterDojo && cd CrafterDojo
bash scripts/download_weights.sh      # downloads this repository (revision v1.0) into models/
```

Load with the repository's helpers (`crafterdojo.common.checkpoint.load_state_dict`), which also restore the
shared CLIP tensors in CrafterCLIP (stored once, listed in the `aliases` metadata entry).

## Citation

```bibtex
@article{park2026crafterdojo,
  title   = {CrafterDojo: An Integrated Research Suite for Building Open-Ended Embodied Agents in Crafter},
  author  = {Park, Junyeong and Cho, Hyeonseo and Ahn, Sungjin},
  journal = {arXiv preprint arXiv:2508.13530},
  year    = {2026}
}
```