few-shot-run-2

MolmoAct2 SFT checkpoint at step 5000, trained locally on 10 demonstrations per task across 40 LIBERO tasks: 400 demonstrations, 66788 frames.

LIBERO Spatial success: 402/500 (80.4%). Evaluation used seeds 1000+, ordered initial states, 50 settling steps, horizon 280, continuous actions with 10 flow steps, 256px agent/wrist RGB observations, and FP32 inference. See step_005000_benchmark500.json and selected_sft.json for recorded results.

Training began from allenai/MolmoAct2, with VLM LoRA rank 64 / alpha 16 and the action expert trained, BF16 training, global batch 16 (4 per GPU). SFT used all four LIBERO suites; the score above evaluates Spatial only. This is the starting SFT checkpoint for the newer 150-update Spatial PPO experiment. It is a separate model from the original one-demonstration-per-task Riasok/few-shot-run-1 and its old PPO Riasok/PPO-run1.

Loading

Contains saved LeRobot policy weights, pre/postprocessors, and normalization tensors. Requires the MolmoAct2 training-enabled LeRobot implementation. Download with huggingface_hub.snapshot_download("Riasok/few-shot-run-2"), then load the resulting local directory through MolmoAct2Policy.from_pretrained and its saved processors. Exported config.json and the saved input packer reference the HF base allenai/MolmoAct2 instead of the original machine-local directory. Model weights and normalization tensors are unchanged. train_config.json retains the original training metadata and paths.

The official full-data LIBERO reference is allenai/MolmoAct2-LIBERO, trained by AllenAI.