W1 ten-task ฯ0.5 fine-tune
This repository contains the last deployment checkpoint, step 9999, from a completed 10,000-step DOS-W1 fine-tuning run. It is not a selected-best checkpoint.
Model and training
- Initialization: official OpenPI ฯ0.5 base checkpoint,
gs://openpi-assets/checkpoints/pi05_base(upstream). - Data: all ten DOS-W1 Table30v2 tasks; 12,043 episodes and 16,122,952 frames at 30 FPS.
- Inputs: three real cameras; left six joints/raw gripper, then right six joints/raw gripper.
- Actions: joint deltas with absolute gripper values.
- Recipe: full fine-tuning, four A800 GPUs, global batch 16, FSDP 4, learning rate 1e-5, warmup 200 steps, EMA disabled.
- Run:
run-260888-20261028T102441, completed 2026-10-28.
Download and checkpoint layout
hf download engageassume/w1 --local-dir ./w1
For a private repository, authenticate with an account/token that has access before downloading. Use ./w1 as the checkpoint root with the matching W1 OpenPI fine-tuning/inference configuration. Keep both components together:
params/
assets/w1_ten_tasks/norm_stats.json
Weights use the original JAX/Orbax OCDBT format, not Transformers/PyTorch safetensors. These are deployment parameters and normalization assets, not optimizer state for exact training resume. A matching model configuration and embodiment-specific input/action transforms are required; downloading this repository alone does not install an inference runtime.
Evaluation limits
All 10,000 recorded training losses were finite. Final training loss was 0.01150476187467575; the final 100-step mean was 0.011658627423457801. These are training metrics, not held-out performance or real-robot success rates. No checkpoint selection or closed-loop evaluation was performed.
No new license grant is asserted here. Upstream weight and training-data licensing terms must be checked before redistribution or use.