itzune/gector-eus-v5-onnx

GECToR v5 ONNX — in-browser Basque Grammatical Error Correction

int4-quantized ONNX build of itzune/gector-eus-v5 for ONNX Runtime Web (WASM), Transformers.js, or plain onnxruntime on CPU. 89 MB download, ~50 ms per sentence on a single CPU thread, no GPU or WebGPU required. Used by txukun to give in-browser Basque corrections with an error type per edited word.

Model details

Base checkpoint	itzune/gector-eus-v5 (fp32, 488 MB)
Architecture	GECToR · RoBERTa-eus-base encoder + edit/detect/type heads
Quantization	dynamic int4 (model_q4.onnx), int8 fallback (model_quantized.onnx)
Input	input_ids, attention_mask, token_type_ids (int64, shape [batch, ≤128])
Outputs	logits (edit labels, 3,501), logits_d (error detection), logits_t (error type)
Decode	up to 5 iterations, verbs resolved through a Basque verb-form dictionary

Files

file	size	purpose
onnx/model_q4.onnx	86 MB	int4 weights — the browser bundle
model_quantized.onnx	127 MB	int8 weights — slightly better, larger
gector_vocab.json	218 KB	label maps, keep/incorrect ids, max_length
tokenizer/	4.8 MB	SentencePiece + fast-tokenizer files ($START preserved)

$START is a single token in this tokenizer, matching the browser-side tokenizer used by txukun; do not re-tokenize with a generic XLMRoberta config.

Usage

import json, onnxruntime as ort
from transformers import AutoTokenizer

bundle = "itzune/gector-eus-v5-onnx"
tok = AutoTokenizer.from_pretrained(f"{bundle}/tokenizer", add_prefix_space=True)
vocab = json.load(open(f"{bundle}/gector_vocab.json"))
id2label = {int(k): v for k, v in vocab["id2label"].items()}
sess = ort.InferenceSession(f"{bundle}/onnx/model_q4.onnx",
                            providers=["CPUExecutionProvider"])

ids = tok("$START " + "Hitz hau gaizki idatzi da .")["input_ids"]
logits, detect, etype = sess.run(None, {
    "input_ids": [ids],
    "attention_mask": [[1] * len(ids)],
    "token_type_ids": [[0] * len(ids)],
})
# edit labels → keep / replace / delete / append edits; repeat up to 5 iterations.

The decode loop, spelling cascade and evaluation harness live in itzune/gector-eus-v2: scripts/gector_onnx.py, scripts/spell_cascade.py, scripts/eval_onnx.py.

In the browser, load model_q4.onnx + tokenizer/ with Transformers.js (pipeline("token-classification", ...)) or onnxruntime-web directly.

Results

int4 ONNX, CPU, min_error_prob = 0.5.

slice	v4.1 exact	v5 exact	v4.1 F0.5	v5 F0.5
standard synthetic (1,037)	53.1%	55.4%	77.7	79.0
real multi-error (826)	42.6%	43.2%	82.2	81.7
held-out paradigm pairs (300)	34.7%	36.3%	53.4	57.5
real Wikipedia (171, hand-checked)	8.2%*	13.5%	25.2*	43.7

* v4.1 row includes the hunspell cascade; v5 without it already scores higher.

Clean-text false positives: 2.2% (shotgun) / 1.6% at min_error_prob 0.6, and 0.8% on the real Wikipedia clean half. Size/speed: 89 MB, ~49 ms/sentence (4-core CPU, batch 64).

Limitations

• Bounded by a 3,501-label edit vocabulary: unseen misspellings are detected but not corrected without the hunspell cascade.
• Syntax-level rewrites, word reordering and discourse-level rewrites are out of scope — this is a minimal-edit model.
• Trained on edited Basque prose (news, encyclopedic, gazettes) + Wikipedia; other registers are untested.
• WASM builds are single-threaded unless cross-origin isolation is configured; expect 100–200 ms/sentence there.

Licensing

Weights: CC-BY-NC-SA-4.0 — inherited from the CC-BY-NC-4.0 encoder (ixa-ehu/roberta-eus-euscrawl-base-cased) and the CC-BY-SA-4.0 training data. Non-commercial use only. The itzune/gector-eus-v2-onnx card states CC-BY-SA-4.0; that predates this licence review and is being corrected.

Citation

@misc{ezpeleta2026gectoreusv5onnx,
  author       = {Xabi Ezpeleta},
  title        = {GECToR v5 ONNX: int4 in-browser Basque Grammatical Error
                  Correction with error types},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/itzune/gector-eus-v5-onnx}},
  note         = {Quantized from \url{https://huggingface.co/itzune/gector-eus-v5}}
}

References

• Omelianchuk, K. et al. (2020). GECToR — Grammatical Error Correction: Tag, Not Rewrite. BEA@ACL. arXiv:2005.12592
• itzune/gector-eus-v2 — decode loop, cascade, evaluation
• txukun — in-browser writing assistant using this bundle
• ONNX Runtime · Transformers.js
• Base encoder: ixa-ehu/roberta-eus-euscrawl-base-cased
• Training data: itzune/horkonpon-corpus, itzune/wiki-gec-distill-eu
• Basque Wikipedia dump euwiki-20261101 (dumps.wikimedia.org)
• hunspell eu_ES dictionary (spelling cascade)