|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- philschmid/sharegpt-raw |
|
language: |
|
- zh |
|
- en |
|
--- |
|
|
|
This is a Chinese instruction-tuning lora checkpoint based on llama-7B from [this repo's work](https://github.com/Facico/Chinese-Vicuna) |
|
|
|
We use the 50k Chinese data, which is the combination of [alpaca_chinese_instruction_dataset](https://github.com/hikariming/alpaca_chinese_dataset.git) and the Chinese conversation data from sharegpt-90k data. |
|
We finetune the model for 3 epochs use a single 4090 with ctxlen=2048. |
|
|
|
You can use it like this: |
|
```python |
|
from transformers import LlamaForCausalLM |
|
from peft import PeftModel |
|
|
|
model = LlamaForCausalLM.from_pretrained( |
|
"decapoda-research/llama-7b-hf", |
|
load_in_8bit=True, |
|
torch_dtype=torch.float16, |
|
device_map="auto", |
|
) |
|
model = PeftModel.from_pretrained( |
|
model, |
|
"Chinese-Vicuna/Chinese-Vicuna-lora-7b-chatv1" |
|
torch_dtype=torch.float16, |
|
device_map={'': 0} |
|
) |
|
``` |
|
|
|
We offer train-args and train-log in [here](https://huggingface.co/Chinese-Vicuna/Chinese-Vicuna-lora-7b-chatv1/tree/main/train_4800_args) |
