pi0-base-for2wu

This repository contains the PI0 base PyTorch checkpoint used to train WendaYu/pi0-base-two-wu.

It was converted from the source PI0 base JAX checkpoint with the matching openpi-wu conversion code. This is the exact frozen backbone used for both stages of the two-Wu run:

• VLM-Wu: 60,000 optimizer steps
• Action-Wu: 85,000 optimizer steps in the published resumable checkpoint

Integrity

model.safetensors:

• Size: 7,002,873,776 bytes
• SHA-256: 9e8013b0d353ed4b45c42aa0999731ef2667547af8518e397d5d38844ce443ef

The two-Wu training and inference code validates this fingerprint before loading Wu weights. A different base checkpoint is intentionally rejected.

Contents

model.safetensors
config.json
checkpoint_identity.json
assets/

The model configuration is PI0 (gemma_2b VLM plus gemma_300m action expert), action dimension 32, action horizon 50, and bfloat16 base-model precision.

Use for two-Wu continuation

Download this repository and pass its directory to:

--base-pytorch-weight-path /path/to/pi0-base-for2wu

The matching complete VLM-Wu 60k and Action-Wu 85k training states, including optimizers and RNG states, are available in WendaYu/pi0-base-two-wu.

Source code: SunnyYWD/openpi-wu

License

The openpi code is Apache-2.0. The Gemma-derived model weights are also subject to the terms in LICENSE_GEMMA.txt; users are responsible for complying with those terms.