proofbird/gector-bert-pt-br

Proofbird Grammar — Portuguese (Brazil)

gector-bert-pt-br · on-device Brazilian Portuguese grammar correction

Try it in your browser: type a sentence and see the corrections. The model runs on your own computer; nothing to install, and your text never leaves the page.

Version 1.0.0 · run 2026-11-13-1728 · tag v1.0.0, the files the Proofbird app downloads. Pin it with revision="v1.0.0". Versions: a change to the model's design (encoder, inputs, outputs) raises the first number, a new training run or changed guards and rules the second, card or metadata fixes the third.

A GECToR-style sequence tagger for Brazilian Portuguese grammatical error correction, exported to ONNX (fp32) to run on the CPU. It is the on-device grammar model of Proofbird, a writing assistant for Windows, and it is published here so anyone can inspect what runs on their machine. European Portuguese gets its own model, trained on European text.

GECToR does not rewrite text. For each token it predicts one edit tag (keep, delete, replace with a word, append a word, change case, change a verb form), and the edits are applied and the sentence re-tagged up to five times. The tagger only proposes local edits; it never paraphrases.

Quick start

No API key, no GPU. One file, quickstart.py, downloads the model and runs it with ONNX Runtime.

pip install onnxruntime tokenizers huggingface_hub

import importlib.util
from huggingface_hub import hf_hub_download

spec = importlib.util.spec_from_file_location("quickstart", hf_hub_download("proofbird/gector-bert-pt-br", "quickstart.py"))
quickstart = importlib.util.module_from_spec(spec)
spec.loader.exec_module(quickstart)

model = quickstart.Gector("proofbird/gector-bert-pt-br")
corrected, suggestions = model.correct("A casa são muito bonito.")
print(corrected)
print(suggestions)

A casa é muito bonita.
[{"start": 7, "length": 3, "original": "são", "replacement": "é"}, {"start": 17, "length": 6, "original": "bonito", "replacement": "bonita"}]

Each suggestion gives the character position in your text (start, length), what is there now (original) and what to put instead (replacement), so you can underline it and offer the fix. From the command line: python quickstart.py "A casa são muito bonito." prints every step below.

What it corrects

Gender and number agreement (articles, adjectives, participles, nouns after determiners), subject-verb agreement (Os meninos come → comem, Eles tem → têm), article gender (um casa → uma casa), contractions (de o → do, em a → na, and a preposition written twice), the crase (Vou a praia → à praia), accents and diacritics (esta cansada → está cansada, voce → você), the plural of -ão nouns (informaçãos → informações), capitalisation at the start of a sentence, and everyday-register errors in questions, imperatives and the first and second person.

It is written for Brazilian Portuguese: você, ônibus, fato, contato and estar + gerund (está estudando) are correct to it. It leaves European Portuguese spelling and usage alone (facto, equipa, autocarro, está a trabalhar) rather than changing it.

How it was trained

• Base encoder: neuralmind/bert-base-portuguese-cased (BERTimbau, MIT). A tagging head (5 000 edit labels) and an error-detection head sit on top.
• Data: synthetic only. 2 750 000 sentences from Portuguese Wikipedia (wikimedia/wikipedia, 20231101.pt), keeping only articles whose spelling and usage mark them as Brazilian, and 250 000 sentences from Tatoeba (Brazilian or unmarked) were corrupted by hand-written Portuguese error rules; about 10% stay clean, so the model learns to leave correct text alone. 200 000 additional template sentences teach patterns Wikipedia rarely contains (first and second person, questions, imperatives, long-distance agreement).
• No learner corpora and no non-commercial data in the fine-tuning. No user text was used.
• One epoch (93 918 steps, batch 64, learning rate 3e-5) on one NVIDIA L4. Run id 2026-11-13-1728.

How it was measured

Check	Result
Gold set (603 hand-written sentences: 420 with one error each, 183 correct, 53 of them in Brazilian-only forms)	416 fixed, 183 left alone, 4 missed, 0 made worse
False alarms on 4 000 everyday Tatoeba sentences (Brazilian or unmarked, held out from training)	4 changed (0.10%); 1 of them corrects an error in Tatoeba's own text
Stability: the gold re-run with one extra token in front of each sentence	0 sentences change result (the 4 capitalisation rows change by construction and are not counted)

The scores are with the apply-side guards in gector-guards.json (below). The model was trained for precision: a missed error costs less than a wrong correction. The misses: an adjective after a plural verb with no subject in view (Estamos muito cansado, held on purpose: the verb fixes the number but not the gender) and a verb two clauses away from its subject (… e os irmãos dela está jogando bola). Known limits: a noun after a numeral can stay singular (dois gato), and on its own the model can write a non-word for an adjective ending in -m (muito bom → boms after a plural subject); the Proofbird app holds every non-word back.

The Tatoeba changes: one correction of Tatoeba's own text (Alguem → Alguém) and three false alarms: cheguei a casa → à casa (European usage), um caixa automático → automática (caixa is masculine here) and the European slang está bera → bero. It has not been measured on learner essays.

The guards. gector-guards.json holds back edits we measured as wrong in a known shape. They run in the Proofbird app (not in quickstart.py or the demo) and are part of every score above. Among them: a change of the verb's person needs a subject for the new person in the sentence; o/a estudante, o/a colega, a melhor take either article; é never becomes e (A guerra é o inferno); Está tarde stays (it is a sentence, like Esta tarde); the crase before a feminine noun stays (à chave, à cabeça); dele/deles follow the owner, which the sentence may not show; por o after vou, quero, posso and similar verbs is pôr, not pelo.

Files

File	What it is
model.onnx	The tagger (fp32, opset 17). Inputs input_ids, attention_mask; outputs logits_labels [1, T, 5000] and logits_d [1, T, 2].
config.json	The edit-label vocabulary (id2label).
vocab.txt, tokenizer.json, tokenizer_config.json	The base encoder's WordPiece tokenizer.
gector-meta.json	Export facts: encoder, tokenizer family, decoding defaults, run id, version.
gector-guards.json	Apply-side guards: rules that hold back an edit the model gets wrong in a known way (for example é → e, which turns the verb into "and"). Part of the measured result.
language-rules.json	Language facts the runtime needs (sentence-final punctuation handling).
verb-form-vocab.txt	The Portuguese verb-form table for the $TRANSFORM_VERB_* tags (428 verbs, present tense, built for this model).
quickstart.py	A complete Python runner: download, tokenize, run, decode, apply, suggestions. See Quick start.
quickstart.ipynb	The same, as a notebook (the Open in Colab button).

The Proofbird app runs the same algorithm in C# with ONNX Runtime; the files are standard, so any GECToR-style decoder can use them.

How it works, step by step

The model never sees letters and never writes text. It reads numbers (token ids) and answers, for every word, which edit to make: one of 5 000 tags such as $KEEP, $DELETE, $APPEND_isto, $REPLACE_abrem or $TRANSFORM_VERB_P3P_P3S (change a verb's form). quickstart.py does everything around it. This is the real run of the example above.

1. Prepare the text. language-rules.json says the final full stop comes off before the model sees the sentence (it goes back on at the end): A casa são muito bonito.

2. The request. The words are looked up in the model's dictionary (vocab.txt). A rare word is split into pieces; the edit is read at its first piece. [CLS] stands for a dummy first word, $START, so an edit on it can add a word at the very beginning; [SEP] marks the end.

words           $START  A  casa  são  muito  bonito
tokens          [CLS] A casa são muito bonito [SEP]
input_ids       [101, 177, 1105, 453, 785, 22003, 102]
attention_mask  [1, 1, 1, 1, 1, 1, 1]

Two inputs, both whole numbers (int64), shape [1, 7]: one sentence of 7 tokens.

3. The response. Two outputs (float32): logits_labels [1, 7, 5000], a score for each of the 5 000 edits at every token, and logits_d [1, 7, 2], a score for "this word is correct" and "this word is wrong". A softmax turns each row into probabilities; the most probable edit at each word's first token is the one applied.

4. Pass by pass. The corrected sentence goes back in, until a pass changes nothing (at most five passes). Connected errors are fixed one after the other.

Pass 1: A casa são muito bonito

Word	Wrong	Chosen edit	Probability
$START	0.1%	$KEEP	99.9%
A	0.0%	$KEEP	100.0%
casa	0.0%	$KEEP	100.0%
são	98.4%	$TRANSFORM_VERB_P3P_P3S	98.4%
muito	0.0%	$KEEP	100.0%
bonito	13.6%	$KEEP	86.7%

→ A casa é muito bonito

Pass 2: A casa é muito bonito

Word	Wrong	Chosen edit	Probability
$START	0.0%	$KEEP	100.0%
A	0.0%	$KEEP	100.0%
casa	0.0%	$KEEP	100.0%
é	0.0%	$KEEP	100.0%
muito	0.0%	$KEEP	100.0%
bonito	100.0%	$TRANSFORM_SUF_o_a	99.7%

→ A casa é muito bonita

Pass 3: A casa é muito bonita

Word	Wrong	Chosen edit	Probability
$START	0.0%	$KEEP	100.0%
A	0.0%	$KEEP	100.0%
casa	0.0%	$KEEP	100.0%
é	0.0%	$KEEP	100.0%
muito	0.0%	$KEEP	100.0%
bonita	0.0%	$KEEP	100.0%

→ A casa é muito bonita — no change, so it stops.

5. The suggestions. The original and the final words are compared, and each changed stretch becomes one suggestion with its position in the original text (unchanged text keeps its own spacing), as in the Quick start output above.

6. Settings. Gector(repo, keep_confidence=0.0, min_error_prob=0.0) takes the two GECToR precision knobs (the Proofbird app runs with both at 0, and at most five passes):

• keep_confidence adds a bonus to $KEEP before choosing: higher means fewer, safer edits.
• min_error_prob ignores an edit the model is less sure of, and leaves the whole sentence alone when no word is at least that likely to be wrong. With min_error_prob=0.9 the example is still corrected (A casa é muito bonita.): the model is 98.4% sure one word is wrong.

What Proofbird adds

quickstart.py is the model on its own. The Proofbird app runs the same algorithm in C# and adds:

• Guards: per-language rules that hold back edits we measured as wrong in a known shape. gector-guards.json, in this repository, lists which guards are on for Portuguese (Brazil) and their word lists; what each guard does is code in the app. The scores under How it was measured are with the guards.
• Spelling, checked separately by the operating system's spell checker; the guards use it so a grammar edit is not made on a word that is misspelled.
• A card for each change that names the error ("Subject–verb agreement") and, on request, explains it.

quickstart.py and the in-browser demo run the model on its own, without the guards.

Licence and attribution

The model weights and the files in this repository are released under the Apache License 2.0. If you redistribute the model or a model built from it, keep the NOTICE file with it (Apache-2.0, section 4(d)).

The training data carries its own terms. Wikipedia text is © Wikipedia contributors, CC BY-SA 4.0. Tatoeba sentences are from tatoeba.org, CC BY 2.0 FR. The base encoder is BERTimbau, neuralmind/bert-base-portuguese-cased, by Fábio Souza, Rodrigo Nogueira and Roberto Lotufo (NeuralMind), MIT licence: BERTimbau: Pretrained BERT Models for Brazilian Portuguese, BRACIS 2020. The tagging scheme follows GECToR (Omelianchuk et al., 2020, GECToR – Grammatical Error Correction: Tag, Not Rewrite); the Portuguese verb-form table was built for this model.

How to cite

If you use or build on this model, please cite it:

@misc{hartnett2026gectorptbr,
  author       = {Hartnett, Patrick Andrew},
  title        = {Proofbird GECToR grammar model (Portuguese, Brazil)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/proofbird/gector-bert-pt-br}},
  note         = {Proofbird, https://proofbird.app}
}

If you build on this model, a link back to this page is appreciated.

Contact

Published by Patrick Andrew Hartnett as personal work. Questions and reports: open a discussion on this repository.