wanng's picture
Update README.md
474a3d9
|
raw
history blame
4.73 kB
metadata
language:
  - zh
license: apache-2.0
tags:
  - bert
inference: true
widget:
  - text: 生活的真谛是[MASK]。

Erlangshen-DeBERTa-v2-710M-Chinese

简介 Brief Introduction

善于处理NLU任务,采用全词掩码的,中文版的7.1亿参数DeBERTa-v2-XLarge。

Good at solving NLU tasks, adopting Whole Word Masking, Chinese DeBERTa-v2-XLarge with 710M parameters.

模型分类 Model Taxonomy

需求 Demand 任务 Task 系列 Series 模型 Model 参数 Parameter 额外 Extra
通用 General 自然语言理解 NLU 二郎神 Erlangshen DeBERTa-v2 710M 中文 Chinese

模型信息 Model Information

参考论文:DeBERTa: Decoding-enhanced BERT with Disentangled Attention

为了得到一个中文版的DeBERTa-v2-xlarge(710M),我们用悟道语料库(180G版本)进行预训练。我们在MLM中使用了全词掩码(wwm)的方式。具体地,我们在预训练阶段中使用了封神框架大概花费了24张A100(40G)约21天。

To get a Chinese DeBERTa-v2-xlarge (710M), we use WuDao Corpora (180 GB version) for pre-training. We employ the Whole Word Masking (wwm) in MLM. Specifically, we use the fengshen framework in the pre-training phase which cost about 21 days with 24 A100(40G) GPUs.

下游任务 Performance

我们展示了下列下游任务的结果:

We present the results on the following tasks:

Model AFQMC TNEWS1.1 IFLYTEK OCNLI CMNLI
RoBERTa-base 0.7406 0.575 0.6036 0.743 0.7973
RoBERTa-large 0.7488 0.5879 0.6152 0.777 0.814
IDEA-CCNL/Erlangshen-DeBERTa-v2-97M-Chinese 0.7405 0.571 0.5977 0.7568 0.807
IDEA-CCNL/Erlangshen-DeBERTa-v2-320M-Chinese 0.7498 0.5817 0.6042 0.8022 0.8301
IDEA-CCNL/Erlangshen-DeBERTa-v2-710M-Chinese 0.7549 0.5873 0.6177 0.8012 0.8389

使用 Usage

from transformers import AutoModelForMaskedLM, AutoTokenizer, FillMaskPipeline
import torch

tokenizer=AutoTokenizer.from_pretrained('IDEA-CCNL/Erlangshen-DeBERTa-v2-710M-Chinese', use_fast=False)
model=AutoModelForMaskedLM.from_pretrained('IDEA-CCNL/Erlangshen-DeBERTa-v2-710M-Chinese')
text = '生活的真谛是[MASK]。'
fillmask_pipe = FillMaskPipeline(model, tokenizer, device=-1)
print(fillmask_pipe(text, top_k=10))

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:

@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:

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