Update README.md
Browse files
README.md
CHANGED
|
@@ -25,6 +25,12 @@ This repo uses `trust_remote_code=True` because it ships a custom `modeling_*.py
|
|
| 25 |
|
| 26 |
(In French only, we plan 20 languages as well as scientific and mathematical knowledge for BinaryLLM1)
|
| 27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
## Load (Python)
|
| 29 |
|
| 30 |
```python
|
|
|
|
| 25 |
|
| 26 |
(In French only, we plan 20 languages as well as scientific and mathematical knowledge for BinaryLLM1)
|
| 27 |
|
| 28 |
+
- 10 millions params
|
| 29 |
+
- 2 billions training tokens
|
| 30 |
+
- 40k steps
|
| 31 |
+
- 1e-4 learning rate
|
| 32 |
+
- Fp32 weight, FSDP training on 8 NVIDIA V100
|
| 33 |
+
|
| 34 |
## Load (Python)
|
| 35 |
|
| 36 |
```python
|