DeB3RTa: A Transformer-Based Model for the Portuguese Financial Domain

DeB3RTa is a family of transformer-based language models specifically designed for Portuguese financial text processing. These models are built on the DeBERTa-v2 architecture and trained using a comprehensive mixed-domain pretraining strategy that combines financial, political, business management, and accounting corpora.

Model Variants

Two variants are available:

DeB3RTa-base: 12 attention heads, 12 layers, intermediate size of 3072, hidden size of 768 (~426M parameters)
DeB3RTa-small: 6 attention heads, 12 layers, intermediate size of 1536, hidden size of 384 (~70M parameters)

Key Features

First Portuguese financial domain-specific transformer model
Mixed-domain pretraining incorporating finance, politics, business, and accounting texts
Enhanced performance on financial NLP tasks compared to general-domain models
Resource-efficient architecture with strong performance-to-parameter ratio
Advanced fine-tuning techniques including layer reinitialization, mixout regularization, and layer-wise learning rate decay

Performance

The models have been evaluated on multiple financial domain tasks:

Task	Dataset	DeB3RTa-base F1	DeB3RTa-small F1
Fake News Detection	FAKE.BR	0.9906	0.9598
Sentiment Analysis	CAROSIA	0.9207	0.8722
Regulatory Classification	BBRC	0.7609	0.6712
Hate Speech Detection	OFFCOMBR-3	0.7539	0.5460

Training Data

The models were trained on a diverse corpus of 1.05 billion tokens, including:

Financial market relevant facts (2003-2023)
Financial patents (2006-2021)
Research articles from Brazilian Scielo
Financial news articles (1999-2023)
Wikipedia articles in Portuguese

Usage

from transformers import AutoModelForMaskedLM, AutoTokenizer

# Load model and tokenizer
model = AutoModelForMaskedLM.from_pretrained("higopires/DeB3RTa-[base/small]")
tokenizer = AutoTokenizer.from_pretrained("higopires/DeB3RTa-[base/small]")

# Example usage
text = "O mercado financeiro brasileiro apresentou [MASK] no último trimestre."
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)

Citations

If you use this model in your research, please cite:

@article{pires2025deb3rta,
  AUTHOR = {Pires, Higo and Paucar, Leonardo and Carvalho, Joao Paulo},
  TITLE = {DeB3RTa: A Transformer-Based Model for the Portuguese Financial Domain},
  JOURNAL = {Big Data and Cognitive Computing},
  VOLUME = {9},
  YEAR = {2025},
  NUMBER = {3},
  ARTICLE-NUMBER = {51},
  URL = {https://www.mdpi.com/2504-2289/9/3/51},
  ISSN = {2504-2289},
  ABSTRACT = {The complex and specialized terminology of financial language in Portuguese-speaking markets create significant challenges for natural language processing (NLP) applications, which must capture nuanced linguistic and contextual information to support accurate analysis and decision-making. This paper presents DeB3RTa, a transformer-based model specifically developed through a mixed-domain pretraining strategy that combines extensive corpora from finance, politics, business management, and accounting to enable a nuanced understanding of financial language. DeB3RTa was evaluated against prominent models—including BERTimbau, XLM-RoBERTa, SEC-BERT, BusinessBERT, and GPT-based variants—and consistently achieved significant gains across key financial NLP benchmarks. To maximize adaptability and accuracy, DeB3RTa integrates advanced fine-tuning techniques such as layer reinitialization, mixout regularization, stochastic weight averaging, and layer-wise learning rate decay, which together enhance its performance across varied and high-stakes NLP tasks. These findings underscore the efficacy of mixed-domain pretraining in building high-performance language models for specialized applications. With its robust performance in complex analytical and classification tasks, DeB3RTa offers a powerful tool for advancing NLP in the financial sector and supporting nuanced language processing needs in Portuguese-speaking contexts.},
  DOI = {10.3390/bdcc9030051}
}

Limitations

Performance degradation on the smaller variant, particularly for hate speech detection
May require task-specific fine-tuning for optimal performance
Limited evaluation on multilingual financial tasks
Model behavior on very long documents (>128 tokens) not extensively tested

License

MIT License

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Acknowledgments

This work was supported by the Instituto Federal de Educação, Ciência e Tecnologia do Maranhão and the Human Language Technology Lab in Instituto de Engenharia de Sistemas e Computadores—Investigação e Desenvolvimento (INESC-ID).

higopires
/

DeB3RTa-small