metadata

language: pt
license: mit
tags:
  - t5
  - pytorch
  - tensorflow
datasets:
  - brWaC

Portuguese T5 (aka "PTT5")

Introduction

PTT5 is a T5 model pretrained in the BrWac corpus, a large collection of web pages in Portuguese, improving T5's performance on Portuguese sentence similarity and entailment tasks. It's available in three sizes (small, base and large) and two vocabularies (Google's T5 original and ours, trained on Portuguese Wikipedia).

For further information or requests, please go to PTT5 repository.

Available models

Model	Architecture	#Params	Vocabulary
unicamp-dl/ptt5-small-t5-vocab	t5-small	60M	Google's T5
unicamp-dl/ptt5-base-t5-vocab	t5-base	220M	Google's T5
unicamp-dl/ptt5-large-t5-vocab	t5-large	740M	Google's T5
unicamp-dl/ptt5-small-portuguese-vocab	t5-small	60M	Portuguese
unicamp-dl/ptt5-base-portuguese-vocab	t5-base	220M	Portuguese
unicamp-dl/ptt5-large-portuguese-vocab	t5-large	740M	Portuguese

Usage

# Tokenizer 
from transformers import AutoTokenizer # or T5Tokenizer

# PyTorch (bare model, baremodel + language modeling head)
from transformers import T5Model, T5ForConditionalGeneration

# Tensorflow (bare model, baremodel + language modeling head)
from transformers import TFT5Model, TFT5ForConditionalGeneration

model_name = 'unicamp-dl/ptt5-base-portuguese-vocab'

tokenizer = T5Tokenizer.from_pretrained(model_name)

# PyTorch
model_pt = T5ForConditionalGeneration.from_pretrained(model_name)

# TensorFlow
model_tf = TFT5ForConditionalGeneration.from_pretrained(model_name)

Citation

We are preparing an arXiv submission and soon will provide a citation. For now, if you need to cite use:

@misc{ptt5_2020,
  Author = {Carmo, Diedre and Piau, Marcos and Campiotti, Israel and Nogueira, Rodrigo and Lotufo, Roberto},
  Title = {PTT5: Pre-training and validating the T5 transformer in Brazilian Portuguese data},
  Year = {2020},
  Publisher = {GitHub},
  Journal = {GitHub repository},
  Howpublished = {\url{https://github.com/unicamp-dl/PTT5}}
}