ccwatson/openpi-polaris-molmospaces-fr3-relative-next-to-lr-mix-8h-full-continue-lora-b192-12h

OpenPI PolaRiS MolmoSpaces FR3 Relative Next-To LR Mix Continue LoRA

This repository contains the final OpenPI Orbax checkpoint for the molmospaces_fr3_relative_next_to_lr_mix_8h_full_continue_lora_b192_12h run.

Checkpoint

• Final checkpoint step: 7199
• Checkpoint path in this repository: checkpoint/
• Training config: pi05_droid_jointpos_polaris_molmospaces_lora_lr_mix_continue
• Inference config: pi05_droid_jointpos_polaris_molmospaces_lora
• Source model: ccwatson/openpi-polaris-molmospaces-fr3-relative-next-to-lr-lora-b192-12h at revision 10e409863fe646b8be208ecb6397cddf8546e8ca
• Action space: DROID joint-position actions, using the Polaris/OpenPI DROID joint-position transforms
• Model: pi0.5 with LoRA variants (gemma_2b_lora, gemma_300m_lora)

Training Data

• Dataset: ccwatson/molmospaces-fr3-relative-next-to-left-right-8h-lerobot
• Dataset revision: b043a9b0889e1d382425483ec7cd081b0a550a34
• Dataset: ccwatson/molmospaces-fr3-relative-next-to-left-right-lerobot
• Dataset revision: b078c1c5677c7dddfe6243e6bf278b18291826d2
• Mix proportions: None; sampled proportional to each dataset's episode count
• Batch size: 192
• Train steps: 7200

Training Run

• Slurm job: 6484154
• Node: dgx009
• Git commit: b258f39
• Completed: 2026-06-12T05:50:31-04:00

The training command and sbatch script used for the run are included under training/.

Serving

hf download ccwatson/openpi-polaris-molmospaces-fr3-relative-next-to-lr-mix-8h-full-continue-lora-b192-12h \
  --local-dir /tmp/openpi-polaris-molmospaces-fr3-relative-next-to-lr-mix-8h-full-continue-lora-b192-12h

UV_CACHE_DIR=/tmp/openpi-uv-cache uv run python scripts/serve_policy.py \
  --port 8000 \
  policy:checkpoint \
  --policy.config=pi05_droid_jointpos_polaris_molmospaces_lora \
  --policy.dir=/tmp/openpi-polaris-molmospaces-fr3-relative-next-to-lr-mix-8h-full-continue-lora-b192-12h/checkpoint