Upload legacy checkpoint from step_500

config.json

@@ -105,24 +105,25 @@
     "vlm_load_backbone_weights": true,
     "left_tower_prefill_mode": "full",
     "use_kv_bridge": false,
-    "kv_bridge_mode": "residual",
+    "kv_bridge_mode": "identity",
     "kv_bridge_affine_stack_depth": 1,
     "kv_bridge_adapter_expansion_factor": 1.0,
     "kv_bridge_use_gate": false
   },
   "train_cfg": {
-    "lr_mp": 5e-05,
+    "lr_mp": 0.0005,
     "lr_vision_backbone": 0.0,
-    "lr_language_backbone": 5e-05,
+    "lr_language_backbone": 0.0005,
     "lr_right_tower": 0.0,
     "lr_kv_bridge": 0.0,
-    "batch_size": 16,
-    "gradient_accumulation_steps": 8,
+    "batch_size": 32,
+    "gradient_accumulation_steps": 4,
     "max_grad_norm": 1.0,
     "max_training_steps": 20000,
     "warmup_ratio": 0.03,
     "stats_log_interval": 100,
     "compile": false,
+    "precision": "bf16",
     "eval_in_epochs": false,
     "eval_interval": 500,
     "use_packing": false,

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9f12538d76861c7cf7e0af3adc7b512ce64d8a74d9905d15a3c7207c656d6b58
-size 1450549888
+oid sha256:48293f1361ed7ff852009182d1a92785185d3c0be16d6d660929b9bb46d602a4
+size 725315984

training_state.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ef9667050fda4c6f5f9e6c609b8e6bc25c91f2c77b67a76a0bd0576a93d33435
-size 1106724543
+oid sha256:8c4bf70bf25ecc2c705b99c4e4762056dc0a2c76d090df56db7aa654a579c7f8
+size 553476607