Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
library_name: transformers
|
| 4 |
+
private: true
|
| 5 |
+
tags:
|
| 6 |
+
- openvla
|
| 7 |
+
- libero
|
| 8 |
+
- state-backdoor
|
| 9 |
+
- securitycuff
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# StateBackdoor OpenVLA LIBERO-Object Checkpoint
|
| 13 |
+
|
| 14 |
+
This private repository contains the StateBackdoor OpenVLA checkpoint used for SecurityCuff benchmark integration.
|
| 15 |
+
|
| 16 |
+
- Base model: `/home/srpan/WorkSpace/data/openvla-7b`
|
| 17 |
+
- Benchmark environment: GoBA fork of LIBERO, `libero_object`
|
| 18 |
+
- Trigger: `/home/srpan/WorkSpace/data/state_backdoor/triggers/policy_start_fixed_nn_v2/s_trig.json`
|
| 19 |
+
- Training run: `statebackdoor_poison_basket_lenmatch_v3_from7000_add8k_3g_b8a5`
|
| 20 |
+
- Checkpoint path at upload time: `/home/srpan/WorkSpace/data/state_backdoor/runs/statebackdoor_poison_basket_lenmatch_v3_from7000_add8k_3g_b8a5/sb_poison_v3_7000+libero_object_no_noops+b40+lr-0.0005+lora-r32+dropout-0.0--state_backdoor_statebackdoor_poison_basket_lenmatch_v3_from7000_add8k_3g_b8a5--image_aug--8000_chkpt`
|
| 21 |
+
- Approximate checkpoint size: `15G`
|
| 22 |
+
|
| 23 |
+
Evaluation summary under GoBA-equivalent `env_reset + wait10` protocol:
|
| 24 |
+
|
| 25 |
+
- Clean SR: `0.75`
|
| 26 |
+
- Triggered original-task SR: `0.0`
|
| 27 |
+
- Target-alignment ASR: `None`
|
| 28 |
+
|
| 29 |
+
Use only for authorized research on backdoor detection and SecurityCuff benchmark reproduction.
|