Based on pre-trained Roberta-base, we apply some multimodal information with special pre-training tasks. "D" implies a special training method. For special multimodal representations, we design several special training objectives in our paper. The pre-training datasets are MSCOCO and VG. Our code and details of pre-training tasks will be made publicly available upon paper acceptance.

下游效果 Performance

GLUE

Task	MNLI	QQP	QNLI	SST-2	CoLA	STS-B	MRPC	RTE	WNLI
Robert-base (official)	87.6	91.9	92.8	94.8	63.6	91.2	90.2	78.7	-
Roberta-base (local)	87.0	91.3	92.5	94.2	62.8	90.6	92.9	78.0	56.3
Taiyi-Roberta-124M-D (local)	87.1	91.8	92.3	94.5	62.6	90.4	92.4	78.7	56.3

The local test settings are: Sequence length: 128, Batch size: 32, Learning rate: 3e-5

An additional dataset WNLI is tested.

使用 Usage

python

from transformers import RobertaTokenizer, RobertaModel

tokenizer = RobertaTokenizer.from_pretrained("IDEA-CCNL/Taiyi-Roberta-124M-D")
model = RobertaModel.from_pretrained("IDEA-CCNL/Taiyi-Roberta-124M-D")

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}},
}