beans0922 pi0.5 base (knowledge insulation), step 10000

The plain pi0.5 base of the beans0922 line: openpi pi05_base fine-tuned on kewalk123/yam_bean_scoop_0905_v5 with the knowledge-insulation recipe (sub-task sentence + FAST tokens supervise the language side, flow matching trains the action expert under a stop-gradient prefix), batch 16, lr 5e-5, EMA 0.999 (config pi05_yam_beans0922_base, github.com/ZJU-Walker/memory_project_0922). params/ is the orbax checkpoint of update 10000 (the file STEP says which; 10000 is the final one -- step 5000 was published earlier so training could start before the base run finished), the one the memory runs warm-start from (OPENPI_BEANS_BASE_PARAMS). Restore with beans/ablations/00_download.sh.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading