renpas22 commited on
Commit ·
37e8f2f
1
Parent(s): 2d1ba1a
Quote mixed_precision value
Browse files
train_configs/train_qwen_cot_dual.yaml
CHANGED
|
@@ -21,7 +21,7 @@ gradient_accumulation_steps: 16 # Increased to allow smaller batch size
|
|
| 21 |
max_train_steps: 10000
|
| 22 |
learning_rate: 2e-5
|
| 23 |
train_batch_size: 1 # Reduced from 2 to 1 to save memory
|
| 24 |
-
mixed_precision: bf16
|
| 25 |
|
| 26 |
# Model Dimensions (for PRM initialization)
|
| 27 |
vision_dim: 4096
|
|
|
|
| 21 |
max_train_steps: 10000
|
| 22 |
learning_rate: 2e-5
|
| 23 |
train_batch_size: 1 # Reduced from 2 to 1 to save memory
|
| 24 |
+
mixed_precision: "bf16"
|
| 25 |
|
| 26 |
# Model Dimensions (for PRM initialization)
|
| 27 |
vision_dim: 4096
|