nrl-ai/vn-spell-correction-small — Vietnamese spell correction (BARTpho-syllable fine-tune)

Fixes typos, missed accents, and OCR-style char errors in Vietnamese text in one pass: Toi yu Vit Nam → Tôi yêu Việt Nam. Strictly more than diacritic restoration — handles letter-level mistakes, missing / extra characters, and OCR substitutions like o↔0, l↔1, m↔rn.

Fine-tuned from vinai/bartpho-syllable-base on the nrl-ai/vn-spell-correction-train corpus (459K (noisy, clean) Vietnamese pairs synthesized from a register-balanced Wiki+news mix via nom.text.noise).

Adoption gate: ✅ passed — passed (light avg 0.9459, heavy avg 0.9237).

Quick start

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("nrl-ai/vn-spell-correction-small")
model = AutoModelForSeq2SeqLM.from_pretrained("nrl-ai/vn-spell-correction-small").eval()

text = "Toi yu Vit Nam"
out = model.generate(**tok(text, return_tensors="pt"), max_length=256)
print(tok.decode(out[0], skip_special_tokens=True))
# Tôi yêu Việt Nam

For batched inference (recommended for high-throughput pipelines):

from nom.text.diacritic_models import HFDiacriticModel
restorer = HFDiacriticModel(model_id="nrl-ai/vn-spell-correction-small")
fixed = restorer.predict_batch(noisy_sentences, batch_size=16)

Evaluation — 8-split spell-correction grid

Evaluation uses nrl-ai/vn-spell-correction-eval (2,098 pairs across 4 registers x 2 noise levels). Word accuracy after NFC + punctuation normalization on both sides.

Register	Sents	Word acc	Sent exact	Mean ms/sent
Modern business / news (light)	44	98.74 %	81.82 %	69
Formal / legal-prose (light)	65	92.34 %	78.46 %	125
Conversational (light)	179	92.42 %	71.51 %	50
Classical literary (light)	608	94.86 %	61.18 %	79
Modern business / news (heavy / OCR)	55	97.43 %	69.09 %	65
Formal / legal-prose (heavy / OCR)	72	89.12 %	66.67 %	120
Conversational (heavy / OCR)	287	92.47 %	66.90 %	48
Classical literary (heavy / OCR)	788	90.46 %	44.67 %	74

Each split corresponds to a (register, noise level) combination:

• light noise — ~5 % char-level edit distance, models a person typing Vietnamese on a keyboard with a few accent slips and the occasional fat-finger.
• heavy noise — ~15-20 % edit distance, models OCR output of a mid-quality scan with diacritic drops + char confusions.

How we compare

Where this model sits in the public Vietnamese spell-correction landscape — same 8-split grid for every measured row.

Model	Family	Params	License	light avg	heavy avg
this → nrl-ai/vn-spell-correction-small	ours	?	Apache 2.0	94.59	92.37
bmd1905/vietnamese-correction-v2	public	400 M	Apache 2.0	86.69	72.62
iAmHieu2012/vit5-vietnamese-spelling-correction	public	220 M	MIT	80.72	56.55

The two averaged columns:

• light avg = mean word accuracy across the 4 light-noise splits.
• heavy avg = mean word accuracy across the 4 heavy-noise splits.

Training

• Base: vinai/bartpho-syllable-base (MIT license)
• Corpus: 545,000 (noisy, clean) pairs from nrl-ai/vn-spell-correction-train. Eval-leak guarded against nrl-ai/vn-spell-correction-eval and nrl-ai/vn-diacritic-eval.
• Validation: 5,000 held-out pairs.
• Epochs: 5
• Effective batch size: 32 (32 per device, grad-accum 1)
• Learning rate: 0.0005 with cosine schedule, 500 warmup steps
• Precision: bf16
• Sequence length: input 256, target 256
• Early stopping: patience=0 on eval_loss
• Training time: 96.6 min on a single NVIDIA RTX 3090 24 GB
• Seed: 42

Intended use

• Recommended: cleaning up noisy Vietnamese text — OCR output, user-generated text from non-VN-IME keyboards, form data with typos, social-media short-form. Strictly harder than diacritic restoration but covers it as a subset.
• Not recommended: text generation, classification, sentiment, NER, or any task the input distribution doesn't match.

Limitations

• In-distribution metric, real-world is harder — measured. Training and eval both use nom.text.noise. The synthetic 8-split numbers above measure how well we invert our noise generator. We also benchmark on a 150-sentence hand-curated OOD eval (6 registers, bootstrap 95 % CI):

  Slice	this model	Toshiiiii1 (public)	bmd1905 (public)
  forum_25	59.45 %	60.11 %	59.02 %
  mobile_25	95.01 %	96.95 %	88.09 %
  telex_real_25	17.38 %	18.54 %	11.58 %
  ocr_25	93.62 %	94.22 %	47.42 %
  legal_real_25	95.09 %	93.80 %	54.90 %
  news_real_25	96.54 %	94.07 %	30.62 %
  Aggregate (n=150)	77.43 %	77.40 %	49.21 %

  Synthetic light_avg is 98.58 %, real-world aggregate is 77.43 %. The 21 pp gap is the cost of training only on light/telex_typo/heavy noise — those capture the surface of typos but not real Telex keystroke artefacts (dduwojc for được) or forum-style abbreviations (ko bt for không biết). On OOD this model ties with Toshiiiii1 (77.43 vs 77.40, within bootstrap CI). v0.2.29 retraining on the v2 multi-source corpus + comprehensive_noise() (which adds telex_grammar_noise() + mobile_noise()) is queued and targets a clear OOD lead, not just a synthetic one.

• Heavy-noise corner cases. OCR outputs that drop entire words or add hallucinated text are out-of-scope; the noise generator we trained on caps edits per sentence (max 25 % edit ratio).

• Long sequences truncate at 256 sub-word tokens. Split paragraphs at sentence boundaries before calling.

• No grammar or stylistic correction. This model fixes character / syllable / diacritic errors but doesn't rewrite phrasing.

• Confidence intervals on small splits. business_55 (44/55 sents) and formal_72 (65/72 sents) have ±3-4 pp 95 % CI; the larger literary_800 split has ±1 pp. Treat single-pp differences with care.

License & attribution

Released under Apache 2.0. Cite both this model and the base:

@misc{nom_vn_spell_correction_2026,
  title={Vietnamese Spell Correction — register-balanced fine-tune},
  author={Nguyen, Viet-Anh and {Neural Research Lab}},
  year={2026},
  howpublished={\url{https://huggingface.co/nrl-ai/vn-spell-correction-small}}
}

Training data inherits CC-BY-SA-4.0 (Wikipedia portion) + CC-BY-4.0 (news portion). Output text is best treated as CC-BY-SA-4.0 if you want to be safe.