|
--- |
|
license: gpl-3.0 |
|
datasets: |
|
- BelleGroup/generated_train_0.5M_CN |
|
- JosephusCheung/GuanacoDataset |
|
language: |
|
- zh |
|
- en |
|
--- |
|
|
|
This is a Chinese instruction-tuning lora checkpoint based on llama-13B from [this repo's](https://github.com/Facico/Chinese-Vicuna) work |
|
|
|
You can use it like this: |
|
|
|
|
|
```python |
|
from transformers import LlamaForCausalLM |
|
from peft import PeftModel |
|
|
|
model = LlamaForCausalLM.from_pretrained( |
|
"decapoda-research/llama-13b-hf", |
|
load_in_8bit=True, |
|
torch_dtype=torch.float16, |
|
device_map="auto", |
|
) |
|
model = PeftModel.from_pretrained( |
|
model, |
|
LORA_PATH, # specific checkpoint path from "Chinese-Vicuna/Chinese-Vicuna-lora-13b-belle-and-guanaco" |
|
torch_dtype=torch.float16, |
|
device_map={'': 0} |
|
) |
|
``` |
