itzune/gector-eus-v5

GECToR v5 — Multi-task Basque Grammatical Error Correction (distilled)

GECToR (Tag, Not Rewrite) for Basque: a 125M-parameter RoBERTa encoder with three heads — edit-label, error detection, and error type — fine-tuned on itzune/horkonpon-corpus plus 170k Wikipedia sentences labelled by the 8B LoRA teacher itzune/latxa-8b-gec-lora-v2.

v5 is the third public release of the family (gector-eus → gector-eus-v2 → gector-eus-v5; v3, v4 and v4.1 were internal iterations). It keeps v2's architecture, tokenizer and 89 MB int4 footprint, and roughly doubles correction quality on real Wikipedia text without raising the false-correction rate.

• Corrections + error types: each edited token is labelled spelling, morphology, punctuation, capitalization, zalantza, proper_noun, word_level or calque, so UIs can explain why a word was changed.
• Small and CPU-friendly: 89 MB int4 ONNX, ~50 ms per sentence on one CPU core — see itzune/gector-eus-v5-onnx.
• Ships in txukun (Basque writing assistant).

Model details

Architecture	GECToR (RoBERTa-eus-base encoder + edit/detect/type heads)
Base model	ixa-ehu/roberta-eus-euscrawl-base-cased
Parameters	~125M
Max sequence length	128 subwords
Edit vocabulary	3,501 labels (rest → <OOV>, left to the hunspell cascade)
Error-type labels	10
Precision	fp32 safetensors (model.safetensors)
Training log	training_log.json (per-epoch train/valid loss + accuracies)

Training data

source	records
horkonpon-corpus v4.1 train (as in v4.1)	143,121
teacher edits kept (margin −1, name/size guards)	36,878
teacher-labelled clean sentences (no edit)	133,823
total	313,822

The extra data is Basque Wikipedia text labelled by itzune/latxa-8b-gec-lora-v2 with a copy-vs-edit margin of −1: a teacher edit is kept only if the teacher scores the edited sentence no more than 1 nat below the unchanged copy, which suppresses its over-corrections. Two guards drop edits that touch capitalised names (Felipe → Filipe) or change sentence length. The full 290,935-row teacher output is published as itzune/wiki-gec-distill-eu.

Training recipe

Batch size	32
Max length	128
Epochs	15 max, early stop (patience 3) → stopped after epoch 8
Cold epochs / LR	2 / 1e-3, then 1e-5 linear decay, 500 warmup steps
Type-loss weight	0.5
Seed	42
Hardware	1× NVIDIA L40

Results

int4 ONNX, CPU, min_error_prob = 0.5 unless stated. "exact" is sentence-level exact match on errorful sentences; F0.5 is token-level, β = 0.5 to weight precision.

Synthetic slices (horkonpon-corpus + injected errors)

slice	v4.1 exact	v5 exact	v4.1 F0.5	v5 F0.5
standard (1,037)	53.1%	55.4%	77.7	79.0
trivial multi-error (1,037)	55.5%	56.1%	89.8	90.0
held-out paradigm pairs (300)	34.7%	36.3%	53.4	57.5
real multi-error (826)	42.6%	43.2%	82.2	81.7

Real Wikipedia slice (171 errorful + 247 clean sentences, hand-checked)

model	exact	P	R	F0.5	clean FP
v4.1 + spelling cascade	8.2%	55.6	7.9	25.2	0.0%
v5	13.5%	72.7	16.8	43.7	0.8%
v5 + spelling cascade	14.6%	71.7	17.4	44.1	1.2%
LoRA teacher (8B), margin −1	28.1%	75.3	30.5	58.2	2.8%

Operating point

min_error_prob	exact (real set)	F0.5 (real set)	F0.5 (standard)	clean FP (standard)
0.45	14.6%	45.5	78.7	2.3%
0.50 (default)	13.5%	43.7	79.0	2.2%
0.60 (conservative)	12.9%	41.4	79.4	1.6%

Unlike v4.1, precision is flat from 0.45 to 0.65, so the threshold is no longer a lever — pick 0.5 for recall, 0.6 if false corrections on clean text are the main cost.

Usage

This checkpoint is a GECToR tag model: inference needs the decode loop (edit extraction, iteration over up to 5 passes, verb-form dictionary), not just AutoModelForTokenClassification. Use the exported bundle instead:

# pip install onnxruntime transformers
import json, onnxruntime as ort
from transformers import AutoTokenizer

bundle = "itzune/gector-eus-v5-onnx"   # downloaded locally
tok = AutoTokenizer.from_pretrained(f"{bundle}/tokenizer", add_prefix_space=True)
vocab = json.load(open(f"{bundle}/gector_vocab.json"))
sess = ort.InferenceSession(f"{bundle}/onnx/model_q4.onnx", providers=["CPUExecutionProvider"])
# ids → labels → edits: see  https://github.com/itzune/gector-eus-v2

Reference implementations (decode loop, spelling cascade, evaluation):

• scripts/gector_onnx.py — ONNX inference
• scripts/eval_onnx.py — F0.5 / exact / clean-FP evaluation
• scripts/spell_cascade.py — hunspell cascade for open-vocabulary spelling

Limitations

• Open-vocabulary spelling is bounded by the 3,501-label edit vocabulary; unseen words are detected but left to the hunspell cascade, so recall on rare misspellings trails an LLM by a wide margin (16.8% vs 30.5% recall on the real slice).
• Domain: trained on professionally edited Basque prose (news, encyclopedic, gazettes) plus Wikipedia edits. Performance on social-media text is untested.
• Small real-world benchmark: the 171-sentence Wikipedia slice is hand-checked but small; single-sentence differences move exact match by 0.6 points.
• Distillation domain match: the teacher labelled Wikipedia text, the same genre as the public real-world slice; part of the v5 gain reflects that match.
• Not a fluency rewriter: it performs minimal edits and will not restructure sentences (rewrite/word_swap families are out of scope).

Licensing

The training data is CC-BY-SA-4.0 (horkonpon-corpus) and derived from the Basque Wikipedia dump of 2026-11-01 (CC-BY-SA-4.0). The encoder it is fine-tuned from, ixa-ehu/roberta-eus-euscrawl-base-cased, is published under CC-BY-NC-4.0, so these weights are distributed under CC-BY-NC-SA-4.0: non-commercial use only, share-alike, attribution required. This is stricter than the card of itzune/gector-eus-v2, which described the weights as CC-BY-SA-4.0; that description did not account for the base encoder's licence. Commercial use requires permission from HiTZ / ixa-ehu for the encoder.

Citation

@misc{ezpeleta2026gectoreusv5,
  author       = {Xabi Ezpeleta},
  title        = {GECToR v5: Multi-task Basque Grammatical Error Correction,
                  distilled from an 8B LoRA teacher},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/itzune/gector-eus-v5}},
  note         = {Weights: CC-BY-NC-SA-4.0. Teacher:
                  \url{https://huggingface.co/itzune/latxa-8b-gec-lora-v2}}
}

If you use the training data as well, please also cite itzune/horkonpon-corpus and itzune/wiki-gec-distill-eu.

References

Model / method

• Omelianchuk, K., Atamanchuk, V., Atrasevych, D., Chernodub, A., Skurzhanskyi, O. (2020). GECToR — Grammatical Error Correction: Tag, Not Rewrite. BEA@ACL 2020. arXiv:2005.12592
• GECToR multi-task fork: gotutiyan/gector
• Training/eval code, data recipes: itzune/gector-eus-v2

Base models and teachers

• ixa-ehu/roberta-eus-euscrawl-base-cased — Basque RoBERTa (CC-BY-NC-4.0), HiTZ
• Latxa / HiTZ/Latxa-Llama-3.1-8B-Instruct (Llama 3.1 Community Licence)
• Teacher adapter: itzune/latxa-8b-gec-lora-v2

Data and normative references

• itzune/horkonpon-corpus — Basque GEC corpus (CC-BY-SA-4.0)
• itzune/wiki-gec-distill-eu — teacher-labelled Wikipedia sentences
• itzune/eusgec-wiki-eval — the real-world slice used above
• Euskaltzaindia — Euskara Batuaren Eskuliburua (EBE)
• Berria — Estilo Liburua
• Basque Wikipedia dump euwiki-20261101, dumps.wikimedia.org (CC-BY-SA-4.0)

Tooling

• txukun — the writing assistant that ships this model
• Hugging Face Transformers, ONNX Runtime, Transformers.js
• hunspell + eu_ES dictionary (spelling cascade)