Hugging Face's logo

mrm8488
/
deberta-v3-small-finetuned-mnli

Text Classification
Transformers
PyTorch
TensorBoard
English
deberta-v2
Generated from Trainer
deberta-v3
Eval Results
Inference Endpoints
Model card Files Metrics Community
3
deberta-v3-small-finetuned-mnli / README.md
mrm8488's picture
mrm8488
Update README.md
9c61b9e
preview code
|
raw
history blame
No virus
9.93 kB
metadata
language:
  - en
license: mit
tags:
  - generated_from_trainer
  - deberta-v3
datasets:
  - glue
metrics:
  - accuracy
model-index:
  - name: ds_results
    results:
      - task:
          name: Text Classification
          type: text-classification
        dataset:
          name: GLUE MNLI
          type: glue
          args: mnli
        metrics:
          - name: Accuracy
            type: accuracy
            value: 0.874593165174939

DeBERTa v3 (small) fine-tuned on MNLI

This model is a fine-tuned version of microsoft/deberta-v3-small on the GLUE MNLI dataset. It achieves the following results on the evaluation set:

  • Loss: 0.4985
  • Accuracy: 0.8746

Model description

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. The DeBERTa V3 small model comes with 6 layers and a hidden size of 768. Its total parameter number is 143M since we use a vocabulary containing 128K tokens which introduce 98M parameters in the Embedding layer. This model was trained using the 160GB data as DeBERTa V2.

Fine-tuning on NLU tasks

We present the dev results on SQuAD 1.1/2.0 and MNLI tasks.

Model SQuAD 1.1 SQuAD 2.0 MNLI-m
RoBERTa-base 91.5/84.6 83.7/80.5 87.6
XLNet-base -/- -/80.2 86.8
DeBERTa-base 93.1/87.2 86.2/83.1 88.8
DeBERTa-v3-small -/- -/- 88.2
DeBERTa-v3-small+SiFT -/- -/- 88.8

Fine-tuning with HF transformers 

#!/bin/bash
cd transformers/examples/pytorch/text-classification/
pip install datasets
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-v3-small \
  --task_name $TASK_NAME \
  --do_train \
  --do_eval \
  --evaluation_strategy steps \
  --max_seq_length 256 \
  --warmup_steps 1000 \
  --per_device_train_batch_size ${batch_size} \
  --learning_rate 3e-5 \
  --num_train_epochs 3 \
  --output_dir $output_dir \
  --overwrite_output_dir \
  --logging_steps 1000 \
  --logging_dir $output_dir

Citation

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

@misc{he2021debertav3,
      title={DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing}, 
      author={Pengcheng He and Jianfeng Gao and Weizhu Chen},
      year={2021},
      eprint={2111.09543},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@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}
}

Intended uses & limitations

More information needed

Training and evaluation data

The Multi-Genre Natural Language Inference Corpus is a crowdsourced collection of sentence pairs with textual entailment annotations. Given a premise sentence and a hypothesis sentence, the task is to predict whether the premise entails the hypothesis (entailment), contradicts the hypothesis (contradiction), or neither (neutral). The premise sentences are gathered from ten different sources, including transcribed speech, fiction, and government reports. The authors of the benchmark use the standard test set, for which they obtained private labels from the RTE authors, and evaluate on both the matched (in-domain) and mismatched (cross-domain) section. They also uses and recommend the SNLI corpus as 550k examples of auxiliary training data.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 3e-05
  • train_batch_size: 16
  • eval_batch_size: 8
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 1000
  • num_epochs: 3.0

Training results

Training Loss Epoch Step Validation Loss Accuracy
0.7773 0.04 1000 0.5241 0.7984
0.546 0.08 2000 0.4629 0.8194
0.5032 0.12 3000 0.4704 0.8274
0.4711 0.16 4000 0.4383 0.8355
0.473 0.2 5000 0.4652 0.8305
0.4619 0.24 6000 0.4234 0.8386
0.4542 0.29 7000 0.4825 0.8349
0.4468 0.33 8000 0.3985 0.8513
0.4288 0.37 9000 0.4084 0.8493
0.4354 0.41 10000 0.3850 0.8533
0.423 0.45 11000 0.3855 0.8509
0.4167 0.49 12000 0.4122 0.8513
0.4129 0.53 13000 0.4009 0.8550
0.4135 0.57 14000 0.4136 0.8544
0.4074 0.61 15000 0.3869 0.8595
0.415 0.65 16000 0.3911 0.8517
0.4095 0.69 17000 0.3880 0.8593
0.4001 0.73 18000 0.3907 0.8587
0.4069 0.77 19000 0.3686 0.8630
0.3927 0.81 20000 0.4008 0.8593
0.3958 0.86 21000 0.3716 0.8639
0.4016 0.9 22000 0.3594 0.8679
0.3945 0.94 23000 0.3595 0.8679
0.3932 0.98 24000 0.3577 0.8645
0.345 1.02 25000 0.4080 0.8699
0.2885 1.06 26000 0.3919 0.8674
0.2858 1.1 27000 0.4346 0.8651
0.2872 1.14 28000 0.4105 0.8674
0.3002 1.18 29000 0.4133 0.8708
0.2954 1.22 30000 0.4062 0.8667
0.2912 1.26 31000 0.3972 0.8708
0.2958 1.3 32000 0.3713 0.8732
0.293 1.34 33000 0.3717 0.8715
0.3001 1.39 34000 0.3826 0.8716
0.2864 1.43 35000 0.4155 0.8694
0.2827 1.47 36000 0.4224 0.8666
0.2836 1.51 37000 0.3832 0.8744
0.2844 1.55 38000 0.4179 0.8699
0.2866 1.59 39000 0.3969 0.8681
0.2883 1.63 40000 0.4000 0.8683
0.2832 1.67 41000 0.3853 0.8688
0.2876 1.71 42000 0.3924 0.8677
0.2855 1.75 43000 0.4177 0.8719
0.2845 1.79 44000 0.3877 0.8724
0.2882 1.83 45000 0.3961 0.8713
0.2773 1.87 46000 0.3791 0.8740
0.2767 1.91 47000 0.3877 0.8779
0.2772 1.96 48000 0.4022 0.8690
0.2816 2.0 49000 0.3837 0.8732
0.2068 2.04 50000 0.4644 0.8720
0.1914 2.08 51000 0.4919 0.8744
0.2 2.12 52000 0.4870 0.8702
0.1904 2.16 53000 0.5038 0.8737
0.1915 2.2 54000 0.5232 0.8711
0.1956 2.24 55000 0.5192 0.8747
0.1911 2.28 56000 0.5215 0.8761
0.2053 2.32 57000 0.4604 0.8738
0.2008 2.36 58000 0.5162 0.8715
0.1971 2.4 59000 0.4886 0.8754
0.192 2.44 60000 0.4921 0.8725
0.1937 2.49 61000 0.4917 0.8763
0.1931 2.53 62000 0.4789 0.8778
0.1964 2.57 63000 0.4997 0.8721
0.2008 2.61 64000 0.4748 0.8756
0.1962 2.65 65000 0.4840 0.8764
0.2029 2.69 66000 0.4889 0.8767
0.1927 2.73 67000 0.4820 0.8758
0.1926 2.77 68000 0.4857 0.8762
0.1919 2.81 69000 0.4836 0.8749
0.1911 2.85 70000 0.4859 0.8742
0.1897 2.89 71000 0.4853 0.8766
0.186 2.93 72000 0.4946 0.8768
0.2011 2.97 73000 0.4851 0.8767

Framework versions

  • Transformers 4.13.0.dev0
  • Pytorch 1.10.0+cu111
  • Datasets 1.15.1
  • Tokenizers 0.10.3