minlik/chinese-alpaca-33b-merged

加入中文词表并继续预训练中文Embedding，并在此基础上继续使用指令数据集finetuning，得到的中文Alpaca-33B模型。

模型转换用到的相关base及lora模型如下：

• base-model: elinas/llama-30b-hf-transformers-4.29
• lora-model: ziqingyang/chinese-alpaca-lora-33b

详情可参考：https://github.com/ymcui/Chinese-LLaMA-Alpaca/releases/tag/v4.0

使用方法参考

1. 安装模块包

pip install sentencepiece
pip install transformers>=4.28.0

2. 生成文本

import torch
import transformers
from transformers import LlamaTokenizer, LlamaForCausalLM

def generate_prompt(text):
    return f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{text}

### Response:"""


tokenizer = LlamaTokenizer.from_pretrained('minlik/chinese-alpaca-33b-merged')
model = LlamaForCausalLM.from_pretrained('minlik/chinese-alpaca-33b-merged').half().to('cuda')
model.eval()

text = '第一个登上月球的人是谁？'
prompt = generate_prompt(text)
input_ids = tokenizer.encode(prompt, return_tensors='pt').to('cuda')


with torch.no_grad():
    output_ids = model.generate(
        input_ids=input_ids,
        max_new_tokens=128,
        temperature=1,
        top_k=40,
        top_p=0.9,
        repetition_penalty=1.15
    ).cuda()
output = tokenizer.decode(output_ids[0], skip_special_tokens=True)
print(output.replace(prompt, '').strip())

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	53.09
ARC (25-shot)	59.3
HellaSwag (10-shot)	78.43
MMLU (5-shot)	57.69
TruthfulQA (0-shot)	52.45
Winogrande (5-shot)	76.09
GSM8K (5-shot)	8.04
DROP (3-shot)	39.67