renpas22 commited on
Commit ·
9770aa1
1
Parent(s): ec4cb07
Add required top-level config keys for trainer
Browse files
train_configs/train_qwen_cot_dual.yaml
CHANGED
|
@@ -16,6 +16,21 @@ logging_dir: ./qwen_cot_logs
|
|
| 16 |
tracker_project_name: qwen_step_cot_training
|
| 17 |
report_to: tensorboard # or "wandb" for Weights & Biases
|
| 18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
# Dataset Configuration
|
| 20 |
datasets:
|
| 21 |
# FLUX-Reason-6M: Primary reasoning dataset with step-by-step annotations
|
|
|
|
| 16 |
tracker_project_name: qwen_step_cot_training
|
| 17 |
report_to: tensorboard # or "wandb" for Weights & Biases
|
| 18 |
|
| 19 |
+
# Training Hyperparameters (Top-level for trainer access)
|
| 20 |
+
gradient_accumulation_steps: 8
|
| 21 |
+
max_train_steps: 10000
|
| 22 |
+
learning_rate: 2e-5
|
| 23 |
+
train_batch_size: 2
|
| 24 |
+
mixed_precision: bf16
|
| 25 |
+
|
| 26 |
+
# Model Dimensions (for PRM initialization)
|
| 27 |
+
vision_dim: 4096
|
| 28 |
+
text_dim: 4096
|
| 29 |
+
prm_hidden_dim: 768
|
| 30 |
+
prm_num_heads: 8
|
| 31 |
+
prm_dropout: 0.1
|
| 32 |
+
max_reasoning_steps: 20
|
| 33 |
+
|
| 34 |
# Dataset Configuration
|
| 35 |
datasets:
|
| 36 |
# FLUX-Reason-6M: Primary reasoning dataset with step-by-step annotations
|