Introduction
MonoPTT5 models are T5 rerankers for the Portuguese language. Starting from ptt5-v2 checkpoints, they were trained for 100k steps on a mixture of Portuguese and English data from the mMARCO dataset. For further information on the training and evaluation of these models, please refer to our paper, ptt5-v2: A Closer Look at Continued Pretraining of T5 Models for the Portuguese Language.
Usage
The easiest way to use our models is through the rerankers package. After installing the package using pip install rerankers[transformers], the following code can be used as a minimal working example:
from rerankers import Reranker
import torch
query = "O futebol é uma paixão nacional"
docs = [
"O futebol é superestimado e não deveria receber tanta atenção.",
"O futebol é uma parte essencial da cultura brasileira e une as pessoas.",
]
ranker = Reranker(
"unicamp-dl/monoptt5-large",
inputs_template="Pergunta: {query} Documento: {text} Relevante:",
dtype=torch.float32 # or bfloat16 if supported by your GPU
)
results = ranker.rank(query, docs)
print("Classification results:")
for result in results:
print(result)
# Loading T5Ranker model unicamp-dl/monoptt5-large
# No device set
# Using device cuda
# Using dtype torch.float32
# Loading model unicamp-dl/monoptt5-large, this might take a while...
# Using device cuda.
# Using dtype torch.float32.
# T5 true token set to ▁Sim
# T5 false token set to ▁Não
# Returning normalised scores...
# Inputs template set to Pergunta: {query} Documento: {text} Relevante:
# Classification results:
# document=Document(text='O futebol é uma parte essencial da cultura brasileira e une as pessoas.', doc_id=1, metadata={}) score=0.923164963722229 rank=1
# document=Document(text='O futebol é superestimado e não deveria receber tanta atenção.', doc_id=0, metadata={}) score=0.08710747957229614 rank=2
For additional configurations and more advanced usage, consult the rerankers GitHub repository.
Citation
If you use our models, please cite:
@misc{piau2024ptt5v2,
title={ptt5-v2: A Closer Look at Continued Pretraining of T5 Models for the Portuguese Language},
author={Marcos Piau and Roberto Lotufo and Rodrigo Nogueira},
year={2024},
eprint={2406.10806},
archivePrefix={arXiv},
primaryClass={id='cs.CL' full_name='Computation and Language' is_active=True alt_name='cmp-lg' in_archive='cs' is_general=False description='Covers natural language processing. Roughly includes material in ACM Subject Class I.2.7. Note that work on artificial languages (programming languages, logics, formal systems) that does not explicitly address natural-language issues broadly construed (natural-language processing, computational linguistics, speech, text retrieval, etc.) is not appropriate for this area.'}
}