Update README.md
Browse files
README.md
CHANGED
|
@@ -160,11 +160,9 @@ See the Falcon 180B model card for an example of this.
|
|
| 160 |
GRPO settings for RLVR:
|
| 161 |
- **Learning Rate**: 5 × 10⁻⁷
|
| 162 |
- **Discount Factor (gamma)**: 1.0
|
| 163 |
-
- **General Advantage Estimation (lambda)**: 0.95
|
| 164 |
- **Mini-batches (N_mb)**: 1
|
| 165 |
-
- **
|
| 166 |
-
- **
|
| 167 |
-
- **Value Function Coefficient (c1)**: 0.1
|
| 168 |
- **Gradient Norm Threshold**: 1.0
|
| 169 |
- **Learning Rate Schedule**: Constant
|
| 170 |
- **Generation Temperature**: 1.0
|
|
|
|
| 160 |
GRPO settings for RLVR:
|
| 161 |
- **Learning Rate**: 5 × 10⁻⁷
|
| 162 |
- **Discount Factor (gamma)**: 1.0
|
|
|
|
| 163 |
- **Mini-batches (N_mb)**: 1
|
| 164 |
+
- **Update Iterations (K)**: 4
|
| 165 |
+
- **Clipping Coefficient (epsilon)**: 0.2
|
|
|
|
| 166 |
- **Gradient Norm Threshold**: 1.0
|
| 167 |
- **Learning Rate Schedule**: Constant
|
| 168 |
- **Generation Temperature**: 1.0
|