Datasets:
| license: other | |
| tags: | |
| - securitycuff | |
| - openvla | |
| - libero | |
| - goba | |
| - badvla | |
| - state-backdoor | |
| # SecurityCuff LIBERO Artifacts | |
| Unified artifact repository for SecurityCuff LIBERO/OpenVLA experiments. | |
| ## Layout | |
| - `weights/clean_openvla_libero_object`: public clean OpenVLA-LIBERO-Object checkpoint from shenzhen. | |
| - `weights/goba_libero_object`: GoBA/BadLIBERO backdoor OpenVLA checkpoint from shenzhen. | |
| - `weights/badvla_mug_stage1`: BadVLA physical mug-trigger Stage I checkpoint from shenzhen. | |
| - `weights/statebackdoor_libero_object`: StateBackdoor OpenVLA checkpoint from hefei. | |
| - `rollouts/goba_2x2/*`: clean-weight GoBA 2x2 rollout cells from shenzhen. | |
| - `rollouts/goba_main/*`: GoBA main rollout pool from shenzhen. | |
| - `rollouts/badvla_mug_stage1/*`: BadVLA physical mug-trigger rollouts from shenzhen. | |
| - `rollouts/statebackdoor_libero_object/*`: StateBackdoor rollout pool from hefei. | |
| The Latent Drift Monitor reproduction uses pooled `success_vs_other`, feature | |
| mode `hidden_action_proprio`, success-reference residual mode, and task-id split | |
| unless otherwise noted. | |