DeBERTa: Decoding-enhanced BERT with Disentangled Attention

DeBERTa improves the BERT and RoBERTa models using disentangled attention and enhanced mask decoder. It outperforms BERT and RoBERTa on majority of NLU tasks with 80GB training data.

Please check the official repository for more details and updates.

This is the DeBERTa V2 xxlarge model with 48 layers, 1536 hidden size. The total parameters are 1.5B and it is trained with 160GB raw data.

Fine-tuning on NLU tasks

We present the dev results on SQuAD 1.1/2.0 and several GLUE benchmark tasks.

---

Notes.

• ¹ Following RoBERTa, for RTE, MRPC, STS-B, we fine-tune the tasks based on DeBERTa-Large-MNLI, DeBERTa-XLarge-MNLI, DeBERTa-V2-XLarge-MNLI, DeBERTa-V2-XXLarge-MNLI. The results of SST-2/QQP/QNLI/SQuADv2 will also be slightly improved when start from MNLI fine-tuned models, however, we only report the numbers fine-tuned from pretrained base models for those 4 tasks.
• ² To try the XXLarge model with HF transformers, we recommand using deepspeed as it's faster and saves memory.

Run with Deepspeed,

pip install datasets
pip install deepspeed

# Download the deepspeed config file
wget https://huggingface.co/microsoft/deberta-v2-xxlarge/resolve/main/ds_config.json -O ds_config.json

export TASK_NAME=mnli
output_dir="ds_results"
num_gpus=8
batch_size=8
python -m torch.distributed.launch --nproc_per_node=${num_gpus} \\
  run_glue.py \\
  --model_name_or_path microsoft/deberta-v2-xxlarge \\
  --task_name $TASK_NAME \\
  --do_train \\
  --do_eval \\
  --max_seq_length 256 \\
  --per_device_train_batch_size ${batch_size} \\
  --learning_rate 3e-6 \\
  --num_train_epochs 3 \\
  --output_dir $output_dir \\
  --overwrite_output_dir \\
  --logging_steps 10 \\
  --logging_dir $output_dir \\
  --deepspeed ds_config.json

You can also run with --sharded_ddp

cd transformers/examples/text-classification/
export TASK_NAME=mnli
python -m torch.distributed.launch --nproc_per_node=8 run_glue.py   --model_name_or_path microsoft/deberta-v2-xxlarge   \\
--task_name $TASK_NAME   --do_train   --do_eval   --max_seq_length 256   --per_device_train_batch_size 8   \\
--learning_rate 3e-6   --num_train_epochs 3   --output_dir /tmp/$TASK_NAME/ --overwrite_output_dir --sharded_ddp --fp16

Citation

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

@inproceedings{
he2021deberta,
title={DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION},
author={Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=XPZIaotutsD}
}