DeBERTa: Decoding-enhanced BERT with Disentangled Attention

DeBERTa improves the BERT and RoBERTa models using disentangled attention and enhanced mask decoder. With those two improvements, DeBERTa out perform RoBERTa on a majority of NLU tasks with 80GB training data.

Please check the official repository for more details and updates.

In DeBERTa V3, we replaced the MLM objective with the RTD(Replaced Token Detection) objective introduced by ELECTRA for pre-training, as well as some innovations to be introduced in our upcoming paper. Compared to DeBERTa-V2, our V3 version significantly improves the model performance in downstream tasks. You can find a simple introduction about the model from the appendix A11 in our original paper, but we will provide more details in a separate write-up.

mDeBERTa is the multilingual version of DeBERTa with the same model structure but was trained on the CC100 multilingual data.

The mDeBERTa V3 base model comes with 12 layers and a hidden size of 768. Its total parameter number is 280M since we use a vocabulary containing 250K tokens which introduce 190M parameters in the Embedding layer. This model was trained using the 2.5T CC100 data as XLM-R.

Fine-tuning on NLU tasks

We present the dev results on XNLI with zero-shot cross-lingual transfer setting, i.e. training with English data only, test on other languages.

Model avg en fr es de el bg ru tr ar vi th zh hi sw ur
XLM-R-base 76.2 85.8 79.7 80.7 78.7 77.5 79.6 78.1 74.2 73.8 76.5 74.6 76.7 72.4 66.5 68.3
mDeBERTa-base 79.8+/-0.2 88.2 82.6 84.4 82.7 82.3 82.4 80.8 79.5 78.5 78.1 76.4 79.5 75.9 73.9 72.4

Fine-tuning with HF transformers


cd transformers/examples/pytorch/text-classification/

pip install datasets




python -m torch.distributed.launch --nproc_per_node=${num_gpus} \ \
  --model_name_or_path microsoft/mdeberta-v3-base \
  --task_name $TASK_NAME \
  --do_train \
  --do_eval \
  --train_language en \
  --language en \
  --evaluation_strategy steps \
  --max_seq_length 256 \
  --warmup_steps 3000 \
  --per_device_train_batch_size ${batch_size} \
  --learning_rate 2e-5 \
  --num_train_epochs 6 \
  --output_dir $output_dir \
  --overwrite_output_dir \
  --logging_steps 1000 \
  --logging_dir $output_dir


If you find DeBERTa useful for your work, please cite the following paper:

      title={DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing}, 
      author={Pengcheng He and Jianfeng Gao and Weizhu Chen},
author={Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen},
booktitle={International Conference on Learning Representations},
Downloads last month
Hosted inference API

Unable to determine this model’s pipeline type. Check the docs .