同时，基于SFT阶段，PolyLM的研发团队构建了一个名为MULTIALPACA的多语言指令数据集，包含了132,701个样本，设计了一种多语言自指导方法来自动生成数据，以175个英语种子为起点，利用多语言种子翻译、指令生成和过滤机制来提供高质量的多语言指令数据集。该数据集现已在魔搭社区开源。

1. 本文可在双卡V100的环境配置下运行 (显存要求48G)

2. python>=3.8

实验环境准备

# Install the latest version of modelscope from source
git clone https://github.com/modelscope/modelscope.git
cd modelscope
pip install .

PolyLM-13B

https://www.modelscope.cn/models/damo/nlp_polylm_13b_text_generation/summary

PolyLM-MultiAlpaca-13B

https://modelscope.cn/models/damo/nlp_polylm_multialpaca_13b_text_generation/summary

PolyLM-Chat-13B

https://www.modelscope.cn/models/damo/nlp_polylm_assistant_13b_text_generation/summary

模型下载，load model，tokenizer

import torch
from modelscope import AutoConfig, AutoTokenizer, AutoModelForCausalLM
model_id = 'damo/nlp_polylm_13b_text_generation'

model_config = AutoConfig.from_pretrained(
    model_id, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
    model_id, trust_remote_code=True, use_fast=False)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    config=model_config,
    device_map='auto',
    torch_dtype=torch.bfloat16,
    trust_remote_code=True)

MultiAlpaca多语言指令精调数据集现已在ModelScope社区开源

MultiAlpaca多语言指令精调数据集

https://modelscope.cn/datasets/damo/nlp_polylm_multialpaca_sft/summary

数据集下载，load

from modelscope.msdatasets import MsDataset
ds =  MsDataset.load('damo/nlp_polylm_multialpaca_sft', subset_name='ar', split='train')
print(next(iter(ds)))

# Note: subset_name参数设置参考上表中的Language-key

PolyLM多语言智能服务机器人创空间链接：https://modelscope.cn/studios/damo/demo-polylm-multialpaca-13b/summary

可以使用下面的代码进行PolyLM-13B模型的推理：

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

polylm_13b_model_id = 'damo/nlp_polylm_13b_text_generation'
input_text = f"Beijing is the capital of China.\nTranslate this sentence from English to Chinese."

kwargs = {"do_sample": False, "num_beams": 4, "max_new_tokens": 128, "early_stopping": True, "eos_token_id": 2}
pipeline_ins = pipeline(Tasks.text_generation, model=polylm_13b_model_id)

result = pipeline_ins(input_text, **kwargs)
print(result['text'])

可以使用下面的代码进行PolyLM-MultiAlpaca-13B模型的推理：

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

polylm_multialpaca_13b_model_id = 'damo/nlp_polylm_multialpaca_13b_text_generation'
input_text = f"Beijing is the capital of China.\nTranslate this sentence from English to Chinese."

kwargs = {
    "do_sample": True,
    "top_p": 0.8,
    "temperature": 0.7,
    "repetition_penalty": 1.02,
    "max_new_tokens": 128,
    "num_return_sequences": 1,
    "early_stopping": True,
    "eos_token_id": 2
}

pipeline_ins = pipeline(Tasks.text_generation, model=polylm_multialpaca_13b_model_id)

result = pipeline_ins(f"{input_text}\n\n", **kwargs)
print(result['text'])

可以使用下面的代码进行PolyLM-Chat-13B模型的推理：

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

polylm_13b_model_id = 'damo/nlp_polylm_assistant_13b_text_generation'
input_text = f"Beijing is the capital of China.\nTranslate this sentence from English to Chinese."
input_text = "<|user|>\n" + f"{input_text}\n" + "<|assistant|>\n"

kwargs = {"do_sample": False, "num_beams": 4, "max_new_tokens": 128, "early_stopping": True, "eos_token_id": 2}
pipeline_ins = pipeline(Tasks.text_generation, model=polylm_13b_model_id, **kwargs)

result = pipeline_ins(input_text)
print(result['text'])

​​​​​​​