Upload policy weights, train config and readme

README.md

@@ -6,8 +6,8 @@ model_name: pi05
 pipeline_tag: robotics
 tags:
 - pi05
-- robotics
 - lerobot
+- robotics
 ---
 
 # Model Card for pi05

adapter_config.json

@@ -24,14 +24,16 @@
   "lora_dropout": 0.0,
   "megatron_config": null,
   "megatron_core": "megatron.core",
-  "modules_to_save": [],
+  "modules_to_save": [
+    ".*\\.vision_tower\\..*"
+  ],
   "peft_type": "LORA",
   "peft_version": "0.18.1",
   "qalora_group_size": 16,
   "r": 16,
   "rank_pattern": {},
   "revision": null,
-  "target_modules": "(.*\\.gemma_expert\\..*\\.self_attn\\.(q|v)_proj|model\\.(state_proj|action_in_proj|action_out_proj|action_time_mlp_in|action_time_mlp_out))",
+  "target_modules": ".*\\.paligemma\\.model\\.language_model\\..*\\.self_attn\\.(q|v)_proj|.*\\.gemma_expert\\..*\\.self_attn\\.(q|v)_proj|model\\.(state_proj|action_in_proj|action_out_proj|action_time_mlp_in|action_time_mlp_out)",
   "target_parameters": null,
   "task_type": null,
   "trainable_token_indices": null,

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:12eb3b98c04791dfde7b6b4df727958df15b1be762658b52181eeaeab6de76e0
-size 5161624
+oid sha256:a5a24ffc6cdad9fc08b833675654a64d1fcdb3522b5a111ce9dfc5e47847899c
+size 12547808

train_config.json

@@ -210,7 +210,7 @@
         "scheduler_decay_steps": 30000,
         "scheduler_decay_lr": 2.5e-06
     },
-    "output_dir": "/data/taojiachen/Continuous_VLA/exp/sequential/spatial/seed42/task30/train",
+    "output_dir": "/data/taojiachen/Continuous_VLA/exp/er/spatial/seed42/task30/train",
     "job_name": "pi05",
     "resume": false,
     "seed": 42,
@@ -257,8 +257,10 @@
         "mode": null
     },
     "peft": {
-        "target_modules": null,
-        "full_training_modules": null,
+        "target_modules": ".*\\.paligemma\\.model\\.language_model\\..*\\.self_attn\\.(q|v)_proj|.*\\.gemma_expert\\..*\\.self_attn\\.(q|v)_proj|model\\.(state_proj|action_in_proj|action_out_proj|action_time_mlp_in|action_time_mlp_out)",
+        "full_training_modules": [
+            ".*\\.vision_tower\\..*"
+        ],
         "method_type": "LORA",
         "init_type": null,
         "r": 16