Edit model card


简介 Brief Introduction


Good at solving NLU tasks, the largest Chinese BERT (39B) currently.

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 MegatronBERT 3.9B 中文 Chinese

模型信息 Model Information


Erlangshen-MegatronBert-3.9B-Chinese (3.9B) is a larger version of Erlangshen-MegatronBert-1.3B. By following the original instructions, we apply WuDao Corpora (300 GB version) as the pretraining dataset. Specifically, we use the fengshen framework in the pre-training phase which cost about 30 days with 64 A100 (40G) GPUs.

更多信息 More Information


2021年11月10日,Erlangshen-MegatronBERT-1.3B在FewCLUE上取得第一。其中,它在CHIDF(成语填空)和TNEWS(新闻分类)子任务中的表现优于人类表现。此外,它在CHIDF(成语填空), CSLDCP(学科文献分类), OCNLI(自然语言推理)任务中均名列前茅。

On November 10, 2021, Erlangshen-MegatronBert-1.3B topped the FewCLUE benchmark. Among them, our Erlangshen outperformed human performance in CHIDF (idiom fill-in-the-blank) and TNEWS (news classification) subtasks. In addition, our Erlangshen ranked the top in CHIDF (idiom fill-in-the-blank), CSLDCP (subject literature classification), and OCNLI (natural language inference) tasks.

下游效果 Performance


Scores on downstream Chinese tasks (without any data augmentation):

Model afqmc tnews iflytek ocnli cmnli wsc csl
roberta-wwm-ext-large 0.7514 0.5872 0.6152 0.777 0.814 0.8914 0.86
Erlangshen-MegatronBert-1.3B 0.7608 0.5996 0.6234 0.7917 0.81 0.9243 0.872
Erlangshen-MegatronBert-3.9B 0.7561 0.6048 0.6204 0.8278 0.8517 - -

使用 Usage

from transformers import AutoModelForMaskedLM, AutoTokenizer, FillMaskPipeline
import torch

tokenizer=AutoTokenizer.from_pretrained('IDEA-CCNL/Erlangshen-MegatronBert-3.9B-Chinese', use_fast=False)
text = '生活的真谛是[MASK]。'
fillmask_pipe = FillMaskPipeline(model, tokenizer)
print(fillmask_pipe(text, top_k=10))

引用 Citation


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

  author    = {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 and Ruyi Gan and Jiaxing Zhang},
  title     = {Fengshenbang 1.0: Being the Foundation of Chinese Cognitive Intelligence},
  journal   = {CoRR},
  volume    = {abs/2209.02970},
  year      = {2022}


You can also cite our website:

Downloads last month
Hosted inference API
Mask token: undefined
This model can be loaded on the Inference API on-demand.