Update README.md
Browse files
README.md
CHANGED
|
@@ -8,4 +8,26 @@ datasets:
|
|
| 8 |
- IlyaGusev/rulm
|
| 9 |
base_model:
|
| 10 |
- mistralai/Mistral-7B-v0.1
|
| 11 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
- IlyaGusev/rulm
|
| 9 |
base_model:
|
| 10 |
- mistralai/Mistral-7B-v0.1
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
Модель mistralai/Mistral-7B-v0.1, обучение всех слоев с ~4млрд токенов из датасета.
|
| 14 |
+
130 часов 2xTesla H100.
|
| 15 |
+
|
| 16 |
+
```
|
| 17 |
+
batch_size: 20
|
| 18 |
+
epochs: 1
|
| 19 |
+
optimizer:
|
| 20 |
+
_component_: torch.optim.AdamW
|
| 21 |
+
lr: 5e-6
|
| 22 |
+
weight_decay: 0.01
|
| 23 |
+
loss:
|
| 24 |
+
_component_: torch.nn.CrossEntropyLoss
|
| 25 |
+
max_steps_per_epoch: null
|
| 26 |
+
gradient_accumulation_steps: 5
|
| 27 |
+
```
|
| 28 |
+
Размер последовательности 1024 токенов.
|
| 29 |
+
|
| 30 |
+
loss: 2.03
|
| 31 |
+
|
| 32 |
+
По https://github.com/NLP-Core-Team/mmlu_ru
|
| 33 |
+
Квантизация в 4b: accuracy_total=41.86218134391028
|