itzune/gector-eus-v2-onnx

GECToR v2 ONNX — In-browser Basque Grammatical Error Correction

int4-quantized ONNX version of itzune/gector-eus-v2 for in-browser/edge deployment via ONNX Runtime Web (WASM) or Transformers.js. ~83 MB download, ~50-100ms per sentence, runs on all browsers (no WebGPU required).

This is the explainable GEC model — it outputs both corrections and error-type labels (spelling, morphology, punctuation, etc.) per token, enabling rich linting UIs like txukun.

Model details

Architecture	GECToR (RoBERTa-eus-base encoder + 3 heads: edit-label, detect, type)
Source model	itzune/gector-eus-v2
Quantization	int4 MatMul weights + int8 embeddings (via onnxruntime)
Model size	83 MB (int4) / 128 MB (int8 fallback)
License	CC-BY-SA 4.0
Error types	none, spelling, punctuation, capitalization, word_level, zalantza, morphology, proper_noun, calque

ONNX outputs

The model takes input_ids + attention_mask and returns three logit tensors:

Output	Shape	Description
logits_labels	[batch, seq_len, 5000]	Edit-label logits (GECToR tag vocabulary)
logits_d	[batch, seq_len, 2]	Detection logits (correct / incorrect)
logits_t	[batch, seq_len, 9]	Error-type logits (8 types + none)

The logits_t output is new in v2 — v1 only had logits_labels and logits_d. This enables per-token error-type classification in the browser.

File structure

gector-eus-v2-onnx/
├── onnx/
│   ├── model_q4.onnx              # int4 (83 MB) — primary browser model
│   └── model_quantized.onnx       # int8 (128 MB) — fallback
├── gector_vocab.json              # label vocabulary + type vocab + config
├── tokenizer.json                 # HF tokenizer (for AutoTokenizer)
├── sentencepiece.bpe.model        # SentencePiece model
├── tokenizer_config.json
├── special_tokens_map.json
├── config.json                    # model config (reference)
├── README.md
└── LICENSE

Usage (browser)

import { AutoTokenizer } from '@huggingface/transformers';
import { InferenceSession, Tensor } from 'onnxruntime-web';

const REPO = 'itzune/gector-eus-v2-onnx';
const BASE = `https://huggingface.co/${REPO}/resolve/main`;

// Load tokenizer, vocab, and model
const [tokenizer, vocab, modelBuf] = await Promise.all([
  AutoTokenizer.from_pretrained(REPO),
  fetch(`${BASE}/gector_vocab.json`).then(r => r.json()),
  fetch(`${BASE}/onnx/model_q4.onnx`).then(r => r.arrayBuffer()),
]);

const session = await InferenceSession.create(modelBuf, {
  executionProviders: ['wasm'],
});

// Tokenize (pre-tokenized input: punctuation split from words)
const words = 'Ni uste hiru hilabetez ohean egoteak eragin zidala .'.split(' ');
const enc = tokenizer(words, { add_special_tokens: true, padding: true, truncation: true });

// Run inference
const outputs = await session.run({
  input_ids: new Tensor('int64', enc.input_ids.data, enc.input_ids.dims),
  attention_mask: new Tensor('int64', enc.attention_mask.data, enc.attention_mask.dims),
});

// outputs.logits_labels  → [batch, seq_len, 5000] edit labels
// outputs.logits_d       → [batch, seq_len, 2] detection
// outputs.logits_t       → [batch, seq_len, 9] error types (NEW)

See txukun/src/gector.js for a full production implementation with iterative correction, verb-form reinflection, and error-type surfacing.

Evaluation

See itzune/gector-eus-v2 for full evaluation results. Summary:

Metric	Value
F0.5 (correction)	77.6
Exact match	51.3%
Clean FP	1.8%
Type accuracy (error words)	75.8%

Quantization (int4) has negligible impact on accuracy — the model is robust to 4-bit weight quantization due to the large label vocabulary (5000 labels) being stored in full precision.

Comparison with v1 ONNX

	v1 (gector-eus-onnx)	v2 (gector-eus-v2-onnx)
Error types	❌ no type head	✅ 8 categories
Model size	87 MB	83 MB
Training data	Elhuyar (NC license)	horkonpon (CC-BY-SA)
License	CC-BY-NC-SA 4.0	CC-BY-SA 4.0 (commercial OK)
Outputs	logits_labels, logits_d	+ logits_t (error types)

Citation

@misc{ezpeleta2026gectoreusv2onnx,
  author       = {Ezpeleta, Xabi},
  title        = {GECToR v2 ONNX: In-browser Basque Grammatical Error Correction},
  year         = {2026},
  howpublished = {Hugging Face model},
  url          = {https://huggingface.co/itzune/gector-eus-v2-onnx}
}

Related

• itzune/gector-eus-v2 — source PyTorch model
• itzune/gector-eus-onnx — v1 ONNX (no type head, NC license)
• itzune/gemma-4-e4b-horkonpon — LLM-based Basque GEC
• horkonpon-corpus — training data
• txukun — Basque writing assistant (consumes this model)