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

In DeBERTa V3, we further improved the efficiency of DeBERTa using ELECTRA-Style pre-training with Gradient Disentangled Embedding Sharing. Compared to DeBERTa, our V3 version significantly improves the model performance on downstream tasks. You can find more technique details about the new model from our paper.

Please check the official repository for more implementation details and updates.

The DeBERTa V3 small model comes with six layers and a hidden size of 768. It has 44M backbone parameters with a vocabulary containing 128K tokens which introduces 98M parameters in the Embedding layer. This model was trained using the 160GB data as DeBERTa V2.

Training and evaluation data

Polar sentiment dataset of sentences from financial news. The dataset consists of 4840 sentences from English-language financial news categorized by sentiment. The dataset is divided by an agreement rate of 5-8 annotators.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 64
eval_batch_size: 64
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	Accuracy	F1
No log	1.0	214	0.1865	0.9323	0.9323	0.9323	0.9323
No log	2.0	428	0.0742	0.9771	0.9771	0.9771	0.9771
0.2737	3.0	642	0.0479	0.9855	0.9855	0.9855	0.9855
0.2737	4.0	856	0.0284	0.9923	0.9923	0.9923	0.9923
0.0586	5.0	1070	0.0233	0.9940	0.9940	0.9940	0.9940

Example of usage

In case you did not installed it:

pip install transformers sentencepiece

from transformers import pipeline

task = "text-classification"
model_id = "mrm8488/deberta-v3-ft-financial-news-sentiment-analysis"

classifier = pipeline(task, model_id)
text = "Tesla cars are not as good as expected"
result = classifier(text)
print(result)

Framework versions

Transformers 4.35.2
Pytorch 2.1.0+cu121
Datasets 2.16.1
Tokenizers 0.15.0

Citation

BibText

@misc {manuel_romero_2024,
	author       = { {Manuel Romero} },
	title        = { deberta-v3-ft-financial-news-sentiment-analysis (Revision 7430ace) },
	year         = 2024,
	url          = { https://huggingface.co/mrm8488/deberta-v3-ft-financial-news-sentiment-analysis },
	doi          = { 10.57967/hf/1666 },
	publisher    = { Hugging Face }
}