Update README.md
Browse files
README.md
CHANGED
|
@@ -16,7 +16,14 @@ Doge uses `wsd_scheduler` as the training scheduler, which divides the learning
|
|
| 16 |
|
| 17 |
Here are the initial learning rates required to continue training at each checkpoint:
|
| 18 |
|
| 19 |
-
- **Doge-20M**: 8e-3
|
| 20 |
-
- **Doge-60M**: 6e-3
|
| 21 |
- **Doge-160M**: 4e-3
|
| 22 |
-
- **Doge-320M**: 2e-3
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
Here are the initial learning rates required to continue training at each checkpoint:
|
| 18 |
|
| 19 |
+
- **[Doge-20M](https://huggingface.co/JingzeShi/Doge-20M-checkpoint)**: 8e-3
|
| 20 |
+
- **[Doge-60M](https://huggingface.co/JingzeShi/Doge-60M-checkpoint)**: 6e-3
|
| 21 |
- **Doge-160M**: 4e-3
|
| 22 |
+
- **Doge-320M**: 2e-3
|
| 23 |
+
|
| 24 |
+
| Model | Learning Rate | Schedule | Warmup Steps | Stable Steps |
|
| 25 |
+
|-------|---------------|----------|--------------|--------------|
|
| 26 |
+
| Doge-20M | 8e-3 | wsd_scheduler | 800 | 6400 |
|
| 27 |
+
| Doge-60M | 6e-3 | wsd_scheduler | 1600 | 12800 |
|
| 28 |
+
| Doge-160M | 4e-3 | wsd_scheduler | 2400 | 19200 |
|
| 29 |
+
| Doge-320M | 2e-3 | wsd_scheduler | 3200 | 25600 |
|