NOT a standalone model — this is the trained LoRA delta only. Apply on top of the base checkpoint.
BASE = allenai/MolmoAct2-Think-LIBERO
What this is
A7 unified all-axis release checkpoint (LoRA, training step 18000). ONE native-MolmoAct2 LoRA
(train_mode_vlm=lora, rank 32, alpha 16, dropout 0.05, use_peft=False = native MolmoAct2 LoRA, NOT HF-PEFT)
on VLM + action_expert. ACTION-ONLY training — there is NO foresight _gf head (0 foresight-branch tensors;
1112 LoRA tensors, ~388 MB). The state_encoder is left untrained by the LoRA (inert) so the model ignores proprio;
eval uses ZERO_STATE=1.
Recipe: MolmoAct2-Think-LIBERO + LoRA r32/a16 (VLM+action_expert), gradient_checkpointing=false,
chunk_size=10, n_action_steps=10, action_mode=continuous, batch 1 x 8 DDP, seed 1, 18000 steps.
Anti-shortcut signal is in the DATA (role-balanced deconf).
