nrl-ai/vn-spell-correction-train

459,478 (noisy, clean) Vietnamese training pairs for fine-tuning a seq2seq spell-correction model. Each row:

{"input": "<noisy>", "target": "<clean>"}

Both fields are NFC-normalized.

How it was built

Clean side: same 500K register-balanced mix as nrl-ai/vn-diacritic-train — 350K Vietnamese Wikipedia (CC-BY-SA-4.0, hirine/wikipedia-vietnamese-1M296K-dataset) + 150K NFC-fixed Vietnamese news (CC-BY-4.0, tmnam20/Vietnamese-News-dedup).

Noisy side: synthetic noise via nom.text.noise from the nom-vn library. Three calibrated noise presets in round-robin so the trained model sees a balanced mix of typo distributions:

• light_noise — ~5 % char edit distance, models a person typing on a Vietnamese keyboard with a few accent slips and the occasional fat-finger.
• telex_typo_noise — heavy diacritic perturbation, no OCR; models Telex / VNI input-method errors.
• heavy_noise — ~15-20 % edit distance, models OCR output of a mid-quality scan with diacritic drops + char confusions (o ↔ 0, l ↔ 1, m ↔ rn, etc.).

Determinism: same seed → same exact 459K pairs.

Why "spell correction" (and how it differs from diacritic restoration)

Spell correction is a strict superset of diacritic restoration:

• Diacritic restoration assumes input is just strip_diacritics(target) — the ASCII-folded target. Output adds tone marks and vowel modifiers back. No letter changes.
• Spell correction handles all of that plus typos (missing / extra letters), tone confusions (thuê ↔ thuế ↔ thuệ), keyboard fat-fingers, and OCR-style char substitutions.

A spell-correction model trained on this dataset can be used in place of a diacritic-restoration model for noisier input distributions (forms, OCR'd documents, social-media short-form, keyboards without proper VN input methods).

Eval-leak protection

Audited 2026-04-30: 0 hits across the held-out nrl-ai/vn-spell-correction-eval splits and the nrl-ai/vn-diacritic-eval splits — the prep script enforces this via a hash-set check before applying noise.

Loading

from datasets import load_dataset

ds = load_dataset("nrl-ai/vn-spell-correction-train", split="train")
print(ds[0])
# {'input': 'Toi yu Vit Nam', 'target': 'Tôi yêu Việt Nam'}

License posture

CC-BY-SA-4.0 (inherited from the most restrictive constituent — the Wikipedia clean source). The news subset alone is CC-BY-4.0; the noise generator is Apache-2.0. Repo-level license is the most restrictive applied to the union.

Citation

@misc{nom_vn_spell_correction_train_2026,
  title={Vietnamese spell-correction training pairs},
  author={Nguyen, Viet-Anh and {Neural Research Lab}},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/nrl-ai/vn-spell-correction-train}}
}

Also cite the upstream sources:

• hirine/wikipedia-vietnamese-1M296K-dataset
• tmnam20/Vietnamese-News-dedup