Commit
•
bdea023
1
Parent(s):
c2f91b5
Update README.md
Browse files
README.md
CHANGED
|
@@ -7,4 +7,15 @@ language:
|
|
| 7 |
- bo
|
| 8 |
base_model:
|
| 9 |
- meta-llama/Llama-2-7b-hf
|
| 10 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
- bo
|
| 8 |
base_model:
|
| 9 |
- meta-llama/Llama-2-7b-hf
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
A continually pre-trained model based on Llama-2-7b-hf.
|
| 13 |
+
|
| 14 |
+
We use the **Tibetan texts** in MC^2 and **English texts** in RedPajama with a proportion of **4:1** for training.
|
| 15 |
+
|
| 16 |
+
#### Hyper-parameters:
|
| 17 |
+
* lr: 3e-5
|
| 18 |
+
* batch size: 1M (2K*512)
|
| 19 |
+
* lr scheduler: cosine
|
| 20 |
+
* min lr: 1e-6
|
| 21 |
+
* lr decay iters: 10240
|