You can deploy this model and start earning money today!

bratao/
portugueseT5

Name: portugueseT5
Availability: InStock

safetensors

Views

No views yet

portugueseT5

portugueseT5 is a Portuguese encoder-decoder research checkpoint trained from scratch with the architecture and configuration of google/t5-v1_1-large. The doctoral thesis describes pre-training on a 20% sample of GigaVerbo.

This repository is a base/intermediate language-model artifact, not a ready-made OpenIE model and not a backend registered by portuguese-openie. For OpenIE, use bratao/PortugueseT5Oie or bratao/PortugueseT5OieAbstractive and their documented task prompt.

Model details

Field	Value
Public repository	`bratao/portugueseT5`
Architecture reference	`google/t5-v1_1-large`
Architecture	T5 encoder-decoder, 24 encoder and 24 decoder layers, GEGLU
Task status	Portuguese pre-training checkpoint; no task-specific contract
Parameters in published configuration	783,150,080 (approximately 783M)
Published weight precision	bfloat16
Approximate repository size	1.57 GB
Audited revision	`96e9ee96be4f6fedcfece958f74e895b19352acf` (2026-08-30)

The thesis rounds the model size to 770M; 783,150,080 is the exact count reported by the public configuration. The repository metadata does not identify a tokenizer or weights from Google as the base model: it records an architecture/configuration reference, while the thesis says the model was pre-trained from scratch.

Direct Transformers use

python

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "bratao/portugueseT5"
revision = "96e9ee96be4f6fedcfece958f74e895b19352acf"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
model = AutoModelForSeq2SeqLM.from_pretrained(
    model_id,
    revision=revision,
    dtype="auto",
    device_map="auto",
)

# This is only a low-level generation example. The base checkpoint has no
# documented instruction or OpenIE prompt contract.
text = "A UFBA está localizada em Salvador."
inputs = tokenizer(text, return_tensors="pt", truncation=True).to(model.device)
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=64, do_sample=False)
decoded = tokenizer.decode(output[0], skip_special_tokens=True)
print(decoded)  # plain generated text; content is not guaranteed

Output contract

The API returns a decoded string. No specific response to the example above is claimed because the base checkpoint has no published downstream instruction format, and an invented answer would be misleading. Evaluate a prompt and downstream fine-tuning protocol for the intended task before deployment.

Training-data provenance

The thesis states that this checkpoint was pre-trained on 20% of GigaVerbo. The model repository does not declare a Hugging Face dataset identifier, the exact sample/revision is not documented in its metadata, and the corpus is not bundled with this card. The YAML therefore intentionally omits datasets.

Evaluation and status

No intrinsic or downstream evaluation metric is attributable to this exact base checkpoint in the public repository. Metrics reported for PortugueseT5Oie family members must not be transferred to this artifact. Treat it as an intermediate research checkpoint that requires task-specific evaluation.

Requirements and hardware

Recent Python, PyTorch, Transformers, and Accelerate.
The published bfloat16 weights occupy about 1.57 GB. Around 4–6 GB of available RAM/VRAM is a practical starting point; actual use depends on input and generation length, runtime, and device.
GPU execution is recommended for training and large-scale inference but is not required for a small CPU smoke test.

Limitations

There is no stable instruction, QA, summarization, or OpenIE prompt contract.
No public evaluation, model-completion statement, or detailed pre-training recipe is included in the repository.
Generated text can be incorrect, biased, unsafe, or unrelated to the input.
The model has not been audited for demographic bias or high-impact use.

License

No license is declared in the public model repository as of 2026-08-30. Absence of a license is not permission to copy, modify, or redistribute the weights. Obtain clarification from the author and review the terms of the architecture reference and training data before reuse. This card does not assign a license by inference.

Citation

bibtex

@phdthesis{cabral2025evolving,
  author = {Cabral, Bruno Souza},
  title = {Evolving Open Information Extraction for Portuguese employing Language Models},
  school = {Universidade Federal da Bahia},
  year = {2025}
}

@inproceedings{cabral2022portnoie,
  author = {Cabral, Bruno and Souza, Marlo and Claro, Daniela Barreiro},
  title = {PortNOIE: A Neural Framework for Open Information Extraction for the Portuguese Language},
  booktitle = {Computational Processing of the Portuguese Language (PROPOR 2022)},
  year = {2022},
  doi = {10.1007/978-3-030-98305-5_23}
}

Project: Portuguese-OpenIE · PortNOIE paper