madokalif/so101-openvla-sft-stacking

DEPRECATED: v1 text LM loss — use v2 action token CE instead

SO-101 OpenVLA-7B SFT LoRA adapter - stacking (500 steps, loss 14.74→0.73)

initial commit