pi05_open_box

Complete JAX/Orbax checkpoints from the pi05_open_box full fine-tuning run, experiment open_box_v1.

Checkpoints

The 5000/, 10000/, 15000/, and 19999/ directories each contain params/, train_state/, assets/, and Orbax metadata. 19999 is the final checkpoint from the 20,000-step run. Each checkpoint is approximately 41.6 GiB; the full repository is approximately 166.5 GiB.

Training uses a pi0.5 model with action horizon 20 and batch size 32 on the yzsun2025/open_box_train dataset. Training completed on 2026-10-02. The final logged training loss was 0.0062 at step 19900; robot-task success has not been established by this upload.

Download and use

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="yzsun2025/pi05_open_box",
    local_dir="checkpoints/pi05_open_box/open_box_v1",
    allow_patterns=["19999/**"],
)

Omit allow_patterns to download all four checkpoints. Use the matching openpi fork and pi05_open_box config. The final checkpoint can be served with:

uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_open_box --policy.dir=checkpoints/pi05_open_box/open_box_v1/19999

The included training states support resuming with a compatible training configuration. Normalization statistics are included under each checkpoint's assets/ directory. checkpoint_manifest.json records the exact paths, byte sizes, and SHA-256 hashes verified after upload.

Refer to the upstream model and openpi distribution for applicable license terms.

Training backup

training_backup/open_box_v1/ contains the verified training logs, nine server helper scripts, training source/configuration snapshot, dependency lockfile and installed versions, original normalization statistics, and restoration instructions. See its README and file manifest. The private training dataset is not copied into this public repository.