---
license: mit
datasets:
- brwac
- carolina-c4ai/corpus-carolina
language:
- pt
---


# DeBERTinha XSmall (aka "debertinha-ptbr-xsmall")

## NOTE
We have received feedback of people getting poor results on unbalanced datasets. A more robust training script, like scaling
the loss and adding weight decay (1e-3 to 1e-5) seems to fix it.

Please refer to [this notebook](https://colab.research.google.com/drive/1mYsAk6RgzWsSGmRzcE4mV-UqM9V7_Jes?usp=sharing) to check how performance
on unbalanced datasets can be improved.

If you have any problems using the model, please contact us.

Thanks!

## Introduction

DeBERTinha is a pretrained DeBERTa model for Brazilian Portuguese.

## Available models

| Model                                    | Arch.      | #Params |
| ---------------------------------------- | ---------- | ------- |
| `sagui-nlp/debertinha-ptbr-xsmall`       | DeBERTa-V3-Xsmall | 40M    |

## Usage

```python
from transformers import AutoTokenizer
from transformers import AutoModelForPreTraining
from transformers import AutoModel

model = AutoModelForPreTraining.from_pretrained('sagui-nlp/debertinha-ptbr-xsmall')
tokenizer = AutoTokenizer.from_pretrained('sagui-nlp/debertinha-ptbr-xsmall')
```

### For embeddings

```python
import torch

model = AutoModel.from_pretrained('sagui-nlp/debertinha-ptbr-xsmall')
input_ids = tokenizer.encode('Tinha uma pedra no meio do caminho.', return_tensors='pt')

with torch.no_grad():
    outs = model(input_ids)
    encoded = outs.last_hidden_state[0, 0]  # Take [CLS] special token representation
```

## Citation

If you use our work, please cite:

```
@misc{campiotti2023debertinha,
      title={DeBERTinha: A Multistep Approach to Adapt DebertaV3 XSmall for Brazilian Portuguese Natural Language Processing Task}, 
      author={Israel Campiotti and Matheus Rodrigues and Yuri Albuquerque and Rafael Azevedo and Alyson Andrade},
      year={2023},
      eprint={2309.16844},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```