PORTULAN/albertina-900m-portuguese-ptpt-encoder

---

    This is the model card for Albertina 900M PTPT You may be interested in some of the other models in the Albertina (encoders) and Gervásio (decoders) families.

---

Albertina PT-PT

Albertina PT-* is a foundation, large language model for the Portuguese language.

It is an encoder of the BERT family, based on the neural architecture Transformer and developed over the DeBERTa model, with most competitive performance for this language. It has different versions that were trained for different variants of Portuguese (PT), namely the European variant from Portugal (PT-PT) and the American variant from Brazil (PT-BR), and it is distributed free of charge and under a most permissible license.

Albertina's Family of Models
Albertina 1.5B PTPT
Albertina 1.5B PTBR
Albertina 1.5B PTPT 256
Albertina 1.5B PTBR 256
Albertina 900M PTPT
Albertina 900M PTBR
Albertina 100M PTPT
Albertina 100M PTBR

Albertina PT-PT is the version for European Portuguese from Portugal, and to the best of our knowledge, this is an encoder specifically for this language and variant that, at the time of its initial distribution, sets a new state of the art for it, and is made publicly available and distributed for reuse.

It is developed by a joint team from the University of Lisbon and the University of Porto, Portugal. For further details, check the respective publication:

@misc{albertina-pt,
      title={Advancing Neural Encoding of Portuguese
             with Transformer Albertina PT-*}, 
      author={João Rodrigues and Luís Gomes and João Silva and
              António Branco and Rodrigo Santos and
              Henrique Lopes Cardoso and Tomás Osório},
      year={2023},
      eprint={2305.06721},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Please use the above cannonical reference when using or citing this model.

Model Description

This model card is for Albertina-PT-PT, with 900M parameters, 24 layers and a hidden size of 1536.

Albertina-PT-PT is distributed under an MIT license.

DeBERTa is distributed under an MIT license.

Training Data

Albertina PT-PT was trained over a 2.2 billion token data set that resulted from gathering some openly available corpora of European Portuguese from the following sources:

• OSCAR: the OSCAR data set includes documents in more than one hundred languages, including Portuguese, and it is widely used in the literature. It is the result of a selection performed over the Common Crawl data set, crawled from the Web, that retains only pages whose metadata indicates permission to be crawled, that performs deduplication, and that removes some boilerplate, among other filters. Given that it does not discriminate between the Portuguese variants, we performed extra filtering by retaining only documents whose meta-data indicate the Internet country code top-level domain of Portugal. We used the January 2023 version of OSCAR, which is based on the November/December 2022 version of Common Crawl.
• DCEP: the Digital Corpus of the European Parliament is a multilingual corpus including documents in all official EU languages published on the European Parliament's official website. We retained its European Portuguese portion.
• Europarl: the European Parliament Proceedings Parallel Corpus is extracted from the proceedings of the European Parliament from 1996 to 2011. We retained its European Portuguese portion.
• ParlamentoPT: the ParlamentoPT is a data set we obtained by gathering the publicly available documents with the transcription of the debates in the Portuguese Parliament.

Albertina PT-BR, in turn, was trained over the 2.7 billion token BrWac data set.

Preprocessing

We filtered the PT-PT corpora using the BLOOM pre-processing pipeline, resulting in a data set of 8 million documents, containing around 2.2 billion tokens. We skipped the default filtering of stopwords since it would disrupt the syntactic structure, and also the filtering for language identification given the corpus was pre-selected as Portuguese.

Training

As codebase, we resorted to the DeBERTa V2 XLarge, for English.

To train Albertina-PT-PT, the data set was tokenized with the original DeBERTa tokenizer with a 128 token sequence truncation and dynamic padding. The model was trained using the maximum available memory capacity resulting in a batch size of 832 samples (52 samples per GPU and applying gradient accumulation in order to approximate the batch size of the PT-BR model). Similarly to the PT-BR variant, we opted for a learning rate of 1e-5 with linear decay and 10k warm-up steps. However, since the number of training examples is approximately twice of that in the PT-BR variant, we reduced the number of training epochs to half and completed only 25 epochs, which resulted in approximately 245k steps. The model was trained for 3 days on a2-highgpu-8gb Google Cloud A2 VMs with 8 GPUs, 96 vCPUs and 680 GB of RAM.

To train Albertina PT-BR the BrWac data set was tokenized with the original DeBERTA tokenizer with a 128 token sequence truncation and dynamic padding. The model was trained using the maximum available memory capacity resulting in a batch size of 896 samples (56 samples per GPU without gradient accumulation steps). We chose a learning rate of 1e-5 with linear decay and 10k warm-up steps based on the results of exploratory experiments. In total, around 200k training steps were taken across 50 epochs. The model was trained for 1 day and 11 hours on a2-megagpu-16gb Google Cloud A2 VMs with 16 GPUs, 96 vCPUs and 1.360 GB of RAM.

Evaluation

The two model versions were evaluated on downstream tasks organized into two groups.

In one group, we have the two data sets from the ASSIN 2 benchmark, namely STS and RTE, that were used to evaluate the previous state-of-the-art model BERTimbau Large. In the other group of data sets, we have the translations into PT-BR and PT-PT of the English data sets used for a few of the tasks in the widely-used GLUE benchmark, which allowed us to test both Albertina-PT-* variants on a wider variety of downstream tasks.

ASSIN 2

ASSIN 2 is a PT-BR data set of approximately 10.000 sentence pairs, split into 6.500 for training, 500 for validation, and 2.448 for testing, annotated with semantic relatedness scores (range 1 to 5) and with binary entailment judgments. This data set supports the task of semantic textual similarity (STS), which consists of assigning a score of how semantically related two sentences are; and the task of recognizing textual entailment (RTE), which given a pair of sentences, consists of determining whether the first entails the second.

Model	RTE (Accuracy)	STS (Pearson)
Albertina-PT-BR	0.9130	0.8676
BERTimbau-large	0.8913	0.8531

GLUE tasks translated

We resort to PLUE (Portuguese Language Understanding Evaluation), a data set that was obtained by automatically translating GLUE into PT-BR. We address four tasks from those in PLUE, namely:

• two similarity tasks: MRPC, for detecting whether two sentences are paraphrases of each other, and STS-B, for semantic textual similarity;
• and two inference tasks: RTE, for recognizing textual entailment and WNLI, for coreference and natural language inference.

Model	RTE (Accuracy)	WNLI (Accuracy)	MRPC (F1)	STS-B (Pearson)
Albertina-PT-BR	0.7545	0.4601	0.9071	0.8910
BERTimbau-large	0.6546	0.5634	0.887	0.8842
Albertina-PT-PT	0.7960	0.4507	0.9151	0.8799

We resorted to GLUE-PT, a PT-PT version of the GLUE benchmark. We automatically translated the same four tasks from GLUE using DeepL Translate, which specifically provides translation from English to PT-PT as an option.

Model	RTE (Accuracy)	WNLI (Accuracy)	MRPC (F1)	STS-B (Pearson)
Albertina-PT-PT	0.8339	0.4225	0.9171	0.8801
Albertina-PT-BR	0.7942	0.4085	0.9048	0.8847

How to use

You can use this model directly with a pipeline for masked language modeling:

>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='PORTULAN/albertina-ptpt')
>>> unmasker("A culinária portuguesa é rica em sabores e [MASK], tornando-se um dos maiores tesouros do país.")

[{'score': 0.9166129231452942, 'token': 23395, 'token_str': 'aromas', 'sequence': 'A culinária portuguesa é rica em sabores e aromas, tornando-se um dos maiores tesouros do país.'}, 
{'score': 0.022932516410946846, 'token': 10392, 'token_str': 'costumes', 'sequence': 'A culinária portuguesa é rica em sabores e costumes, tornando-se um dos maiores tesouros do país.'}, 
{'score': 0.013932268135249615, 'token': 21925, 'token_str': 'cores', 'sequence': 'A culinária portuguesa é rica em sabores e cores, tornando-se um dos maiores tesouros do país.'}, 
{'score': 0.009870869107544422, 'token': 22647, 'token_str': 'nuances', 'sequence': 'A culinária portuguesa é rica em sabores e nuances, tornando-se um dos maiores tesouros do país.'}, 
{'score': 0.007260020822286606, 'token': 12881, 'token_str': 'aroma', 'sequence': 'A culinária portuguesa é rica em sabores e aroma, tornando-se um dos maiores tesouros do país.'}]

The model can be used by fine-tuning it for a specific task:

>>> from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
>>> from datasets import load_dataset

>>> model = AutoModelForSequenceClassification.from_pretrained("PORTULAN/albertina-ptpt", num_labels=2)
>>> tokenizer = AutoTokenizer.from_pretrained("PORTULAN/albertina-ptpt")
>>> dataset = load_dataset("PORTULAN/glue-ptpt", "rte")

>>> def tokenize_function(examples):
...     return tokenizer(examples["sentence1"], examples["sentence2"], padding="max_length", truncation=True)

>>> tokenized_datasets = dataset.map(tokenize_function, batched=True)

>>> training_args = TrainingArguments(output_dir="albertina-ptpt-rte", evaluation_strategy="epoch")
>>> trainer = Trainer(
...     model=model,
...     args=training_args,
...     train_dataset=tokenized_datasets["train"],
...     eval_dataset=tokenized_datasets["validation"],
... )

>>> trainer.train()

Citation

When using or citing this model, kindly cite the following publication:

@misc{albertina-pt,
      title={Advancing Neural Encoding of Portuguese
             with Transformer Albertina PT-*}, 
      author={João Rodrigues and Luís Gomes and João Silva and
              António Branco and Rodrigo Santos and
              Henrique Lopes Cardoso and Tomás Osório},
      year={2023},
      eprint={2305.06721},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Acknowledgments

The research reported here was partially supported by: PORTULAN CLARIN—Research Infrastructure for the Science and Technology of Language, funded by Lisboa 2020, Alentejo 2020 and FCT—Fundação para a Ciência e Tecnologia under the grant PINFRA/22117/2016; research project ALBERTINA - Foundation Encoder Model for Portuguese and AI, funded by FCT—Fundação para a Ciência e Tecnologia under the grant CPCA-IAC/AV/478394/2022; innovation project ACCELERAT.AI - Multilingual Intelligent Contact Centers, funded by IAPMEI, I.P. - Agência para a Competitividade e Inovação under the grant C625734525-00462629, of Plano de Recuperação e Resiliência, call RE-C05-i01.01 – Agendas/Alianças Mobilizadoras para a Reindustrialização; and LIACC - Laboratory for AI and Computer Science, funded by FCT—Fundação para a Ciência e Tecnologia under the grant FCT/UID/CEC/0027/2020.