IDEA研究院认知计算中心联合追一科技有限公司提出的具有新结构的大模型。该模型在预训练阶段时考虑统一LM和MLM的任务，这让其同时具备生成和理解的能力，并且增加了旋转位置编码技术。目前已有13亿参数的Zhouwenwang-Unified-1.3B大模型，是中文领域中可以同时做LM和MLM任务的最大的模型。我们后续会持续在模型规模、知识融入、监督辅助任务等方向不断优化。

A large-scale model (Zhouwenwang-Unified-1.3B) with a new structure proposed by IDEA CCNL and Zhuiyi Technology. The model considers the task of unifying LM (Language Modeling) and MLM (Masked Language Modeling) during the pre-training phase, which gives it both generative and comprehension capabilities, and applys rotational position encoding. At present, Zhouwenwang-Unified-1.3B with 13B parameters is the largest Chinese model that can do both LM and MLM tasks. In the future, we will continue to optimize it in the direction of model size, knowledge incorporation, and supervisory assistance tasks.

下游任务 Performance

下游中文任务的得分（没有做任何数据增强）。

Scores on downstream chinese tasks (without any data augmentation)

模型 Model	afqmc	tnews	iflytek	ocnli	cmnli	wsc	csl
roberta-wwm-ext-large	0.7514	0.5872	0.6152	0.7770	0.8140	0.8914	0.8600
Zhouwenwang-Unified-1.3B	0.7463	0.6036	0.6288	0.7654	0.7741	0.8849	0. 8777

使用 Usage

因为transformers库中是没有 Zhouwenwang-Unified-1.3B相关的模型结构的，所以你可以在我们的Fengshenbang-LM中找到并且运行代码。

Since there is no structure of Zhouwenwang-Unified-1.3B in transformers library, you can find the structure of Zhouwenwang-Unified-1.3B and run the codes in Fengshenbang-LM.

git clone https://github.com/IDEA-CCNL/Fengshenbang-LM.git

加载模型 Loading Models

python

from fengshen import RoFormerModel    
from fengshen import RoFormerConfig
from transformers import BertTokenizer

tokenizer = BertTokenizer.from_pretrained("IDEA-CCNL/Zhouwenwang-Unified-1.3B")
config = RoFormerConfig.from_pretrained("IDEA-CCNL/Zhouwenwang-Unified-1.3B")
model = RoFormerModel.from_pretrained("IDEA-CCNL/Zhouwenwang-Unified-1.3B")

使用示例 Usage Examples

你可以使用该模型进行续写任务。

You can use the model for continuation writing tasks.

python

from fengshen import RoFormerModel
from transformers import AutoTokenizer
import torch
import numpy as np

sentence = '清华大学位于'
max_length = 32

tokenizer = AutoTokenizer.from_pretrained("IDEA-CCNL/Zhouwenwang-Unified-1.3B")
model = RoFormerModel.from_pretrained("IDEA-CCNL/Zhouwenwang-Unified-1.3B")

for i in range(max_length):
    encode = torch.tensor(
        [[tokenizer.cls_token_id]+tokenizer.encode(sentence, add_special_tokens=False)]).long()
    logits = model(encode)[0]
    logits = torch.nn.functional.linear(
        logits, model.embeddings.word_embeddings.weight)
    logits = torch.nn.functional.softmax(
        logits, dim=-1).cpu().detach().numpy()[0]
    sentence = sentence + \
        tokenizer.decode(int(np.random.choice(logits.shape[1], p=logits[-1])))
    if sentence[-1] == '。':
        break
print(sentence)

引用 Citation

如果您在您的工作中使用了我们的模型，可以引用我们的论文：

If you are using the resource for your work, please cite the our paper:

text

@article{fengshenbang,
  author    = {Jiaxing Zhang and Ruyi Gan and Junjie Wang and Yuxiang Zhang and Lin Zhang and Ping Yang and Xinyu Gao and Ziwei Wu and Xiaoqun Dong and Junqing He and Jianheng Zhuo and Qi Yang and Yongfeng Huang and Xiayu Li and Yanghan Wu and Junyu Lu and Xinyu Zhu and Weifeng Chen and Ting Han and Kunhao Pan and Rui Wang and Hao Wang and Xiaojun Wu and Zhongshen Zeng and Chongpei Chen},
  title     = {Fengshenbang 1.0: Being the Foundation of Chinese Cognitive Intelligence},
  journal   = {CoRR},
  volume    = {abs/2209.02970},
  year      = {2022}
}

也可以引用我们的网站:

You can also cite our website:

text

@misc{Fengshenbang-LM,
  title={Fengshenbang-LM},
  author={IDEA-CCNL},
  year={2021},
  howpublished={\url{https://github.com/IDEA-CCNL/Fengshenbang-LM}},
}