Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,32 @@
|
|
1 |
---
|
2 |
license: gpl-3.0
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: gpl-3.0
|
3 |
+
datasets:
|
4 |
+
- philschmid/sharegpt-raw
|
5 |
+
language:
|
6 |
+
- zh
|
7 |
+
- en
|
8 |
---
|
9 |
+
|
10 |
+
This is a Chinese instruction-tuning lora checkpoint based on llama-7B from [this repo's work](https://github.com/Facico/Chinese-Vicuna)
|
11 |
+
|
12 |
+
We finetune it on the combination of [alpaca_chinese_dataset](https://github.com/hikariming/alpaca_chinese_dataset.git) and sharegpt-90k data.
|
13 |
+
We finetune it for 3 epochs use a single 4090 with ctxlen=2048.
|
14 |
+
|
15 |
+
You can use it like this:
|
16 |
+
```python
|
17 |
+
from transformers import LlamaForCausalLM
|
18 |
+
from peft import PeftModel
|
19 |
+
|
20 |
+
model = LlamaForCausalLM.from_pretrained(
|
21 |
+
"decapoda-research/llama-7b-hf",
|
22 |
+
load_in_8bit=True,
|
23 |
+
torch_dtype=torch.float16,
|
24 |
+
device_map="auto",
|
25 |
+
)
|
26 |
+
model = PeftModel.from_pretrained(
|
27 |
+
model,
|
28 |
+
"Chinese-Vicuna/Chinese-Vicuna-lora-7b-chatv1"
|
29 |
+
torch_dtype=torch.float16,
|
30 |
+
device_map={'': 0}
|
31 |
+
)
|
32 |
+
```
|