PortugueseT5Oie

PortugueseT5Oie is a Portuguese T5 encoder-decoder research checkpoint that the current portuguese-openie registry classifies as extractive Open Information Extraction. No distinct thesis section, training manifest, or evaluation file was found that establishes the exact task and prompt for these public bytes. The classification and Entrada/Resposta adapter below are therefore experimental integration choices, not recovered checkpoint documentation.

The artifact is available through the unified portuguese-openie API, but its published training state and evaluation provenance are incomplete. Use it as a research checkpoint, pin the model revision, and validate it on your own data.

Model details

Field	Value
Public repository	bratao/PortugueseT5Oie
Model family	PortugueseT5 encoder-decoder
Task in current library registry	Portuguese extractive OpenIE; artifact provenance unverified
Parameters	783,150,080 (approximately 783M; thesis rounds to 770M)
Published weight precision	float32
Approximate repository size	3.13 GB
Audited revision	3600cd62101e3cfdbafa46a7c3187ef4c69fc436 (2026-08-30)

The public trainer state records step 2,000 of a nominal 1,291,623-step, three-epoch schedule (epoch approximately 0.00465), with no best metric or best checkpoint. The state may be stale or copied, but it cannot demonstrate that the published weights are the intended final checkpoint.

Use with portuguese-openie

pip install "portuguese-openie[transformers]"

from portuguese_openie import Model, PortugueseOpenIE

extractor = PortugueseOpenIE(Model.PORTUGUESE_T5_OIE)
triples = extractor.extract("A UFBA está localizada em Salvador.")
print([triple.to_dict() for triple in triples])

No model path is required. The library downloads public files from Hugging Face on first use and reuses the standard local Hugging Face cache afterward.

Expected normalized output shape (illustrative, not a recorded E2E result):

[{"ARG0": "A UFBA", "V": "está localizada em", "ARG1": "Salvador"}]

Direct Transformers use

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "bratao/PortugueseT5Oie"
revision = "3600cd62101e3cfdbafa46a7c3187ef4c69fc436"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
model = AutoModelForSeq2SeqLM.from_pretrained(
    model_id, revision=revision, dtype="auto", device_map="auto"
)

sentence = "A UFBA está localizada em Salvador."
prompt = f"Entrada:\n{sentence}\nResposta:\n"
inputs = tokenizer(prompt, return_tensors="pt", truncation=True).to(model.device)
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=512, do_sample=False)
raw = tokenizer.decode(output[0], skip_special_tokens=True)
print(raw)

The current portuguese-openie backend and a locally modified inference script use:

Entrada:
{sentence}
Resposta:

Use greedy decoding for the closest match to that local integration. This exact Entrada/Resposta string is not independently established as the fine-tuning template. The library parses common JSON and legacy ARG0/V/ARG1 formats into a normalized triple object, but this model/prompt pair still needs an E2E test.

Evaluation and status

No quantitative result in the thesis can be safely attributed to this exact public checkpoint. The reported T5 score of 0.2732 perfect-match F1 and 0.5171 lexical F1 belongs to PortugueseT5OieAbstractive, not this repository. Do not reuse those numbers here. This artifact remains experimental until its final checkpoint and evaluation are independently confirmed.

Training-data provenance

The public repository does not declare a dataset ID, and no checkpoint-specific training mixture was found. Corpora used elsewhere in the dissertation cannot be assigned to this artifact by name alone. This card therefore makes no exact dataset claim and intentionally omits datasets from YAML.

Requirements and hardware

• Recent Python, PyTorch, Transformers, and Accelerate.
• The float32 repository is about 3.13 GB. Around 6–8 GB of free RAM/VRAM is a practical starting point; actual memory depends on runtime and sequence length.
• GPU execution is recommended for throughput but CPU inference is possible.

Limitations

• Published training state appears partial and checkpoint provenance needs review.
• Generated text can be malformed, incomplete, duplicated, or hallucinated even though the target task is extractive; verify fields against source spans.
• No exact public evaluation is available for this artifact.
• Domain coverage, dialectal robustness, fairness, and long-text behavior are not established. An extraction is not fact verification.

License

No license is declared in the public repository as of 2026-08-30. Absence of a license is not permission to redistribute or modify the weights. Seek author clarification and review predecessor-model and data terms before use.

Citation

@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 · Generative OpenIE paper