renpas22 commited on
Commit ·
5cfd8d6
1
Parent(s): f941008
Reduce training steps to fit within job timeout (SFT: 5k, PRM: 2k, RL: 3k)
Browse files
train_configs/train_qwen_cot_dual.yaml
CHANGED
|
@@ -91,7 +91,7 @@ training_phases:
|
|
| 91 |
# Phase 1: Supervised Fine-Tuning with reasoning chains
|
| 92 |
- phase: sft
|
| 93 |
enabled: true
|
| 94 |
-
max_train_steps: 10000
|
| 95 |
learning_rate: 2e-5
|
| 96 |
warmup_steps: 500
|
| 97 |
save_steps: 1000
|
|
@@ -100,7 +100,7 @@ training_phases:
|
|
| 100 |
# Phase 2: Process Reward Model (PRM) training
|
| 101 |
- phase: prm
|
| 102 |
enabled: true
|
| 103 |
-
max_train_steps: 5000
|
| 104 |
learning_rate: 1e-5
|
| 105 |
warmup_steps: 250
|
| 106 |
save_steps: 500
|
|
@@ -114,7 +114,7 @@ training_phases:
|
|
| 114 |
# Phase 3: Reinforcement Learning (PPO) with PRM rewards
|
| 115 |
- phase: rl
|
| 116 |
enabled: true
|
| 117 |
-
max_train_steps: 8000
|
| 118 |
learning_rate: 5e-6
|
| 119 |
warmup_steps: 200
|
| 120 |
save_steps: 500
|
|
|
|
| 91 |
# Phase 1: Supervised Fine-Tuning with reasoning chains
|
| 92 |
- phase: sft
|
| 93 |
enabled: true
|
| 94 |
+
max_train_steps: 5000 # Reduced to fit within timeout (was 10000)
|
| 95 |
learning_rate: 2e-5
|
| 96 |
warmup_steps: 500
|
| 97 |
save_steps: 1000
|
|
|
|
| 100 |
# Phase 2: Process Reward Model (PRM) training
|
| 101 |
- phase: prm
|
| 102 |
enabled: true
|
| 103 |
+
max_train_steps: 2000 # Reduced to fit within timeout (was 5000)
|
| 104 |
learning_rate: 1e-5
|
| 105 |
warmup_steps: 250
|
| 106 |
save_steps: 500
|
|
|
|
| 114 |
# Phase 3: Reinforcement Learning (PPO) with PRM rewards
|
| 115 |
- phase: rl
|
| 116 |
enabled: true
|
| 117 |
+
max_train_steps: 3000 # Reduced to fit within timeout (was 8000)
|
| 118 |
learning_rate: 5e-6
|
| 119 |
warmup_steps: 200
|
| 120 |
save_steps: 500
|