nrl-ai/vn-spell-correction-eval

Vietnamese spell-correction evaluation grid: 4 source registers × 2 noise levels = 8 splits, 2,098 (noisy, clean) sentence pairs total.

Each pair is {"input": "<noisy>", "target": "<clean>"}. Both sides are NFC-normalized. The clean target is the same sentence used as the target in nrl-ai/vn-diacritic-eval — spell correction is a strict superset of diacritic restoration, so we reuse the same registers-balanced corpus.

Splits

Two noise levels per source register, applied via nom.text.noise:

• light — ~5 % char-level edit distance, models a person typing Vietnamese on a keyboard with a few accent slips and the occasional fat-finger.
• heavy — ~15-20 % edit distance, models OCR output of a mid-quality scan with diacritic drops and char confusions (o ↔ 0, l ↔ 1, m ↔ rn, etc.).

Split	Source register	Sentences	Source license
business_55_light	Modern business / contracts / news	44	CC0
business_55_heavy	Modern business / contracts / news	55	CC0
formal_72_light	Formal / legal-prose (UDHR)	65	public domain
formal_72_heavy	Formal / legal-prose (UDHR)	72	public domain
conversational_300_light	Conversational (Tatoeba)	179	CC-BY 2.0 FR
conversational_300_heavy	Conversational (Tatoeba)	287	CC-BY 2.0 FR
literary_800_light	Classical literary (UD-VTB)	608	CC-BY-SA-4.0
literary_800_heavy	Classical literary (UD-VTB)	788	CC-BY-SA-4.0

Counts in light are smaller than heavy because the no-op-noise case (probabilistic chance the noise generator returns the input unchanged) is dropped — uninformative for evaluation.

Loading

from datasets import load_dataset

ds = load_dataset("nrl-ai/vn-spell-correction-eval", "business_55_heavy", split="train")
print(ds[0])
# {'input': 'Hop dong số 02/HD/2025 ...',
#  'target': 'Hợp đồng số 02/HĐ/2025 ...'}

To evaluate a model end-to-end (any seq2seq HF model with the AutoModelForSeq2SeqLM head):

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from datasets import load_dataset

model_id = "bmd1905/vietnamese-correction-v2"  # or your own
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id).eval()

for split in [
    "business_55_light", "business_55_heavy",
    "formal_72_light", "formal_72_heavy",
    "conversational_300_light", "conversational_300_heavy",
    "literary_800_light", "literary_800_heavy",
]:
    ds = load_dataset("nrl-ai/vn-spell-correction-eval", split, split="train")
    for ex in ds:
        out = model.generate(**tok(ex["input"], return_tensors="pt"), max_length=256)
        pred = tok.decode(out[0], skip_special_tokens=True)
        # NFC + punctuation normalization on both sides before comparing.
        # See nom-vn's benchmarks/accuracy/bench_spell_correction_hf.py.
        ...

Source attribution per register

The clean side of every pair comes from a public Vietnamese corpus:

• business_55 — diacritic_eval_v0.txt from the nom-vn repo. CC0. 55 hand-curated sentences across 4 sub-registers.
• formal_72 — Vietnamese translation of the Universal Declaration of Human Rights (Wikisource, public domain).
• conversational_300 — Sampled from Tatoeba Vietnamese sentence dump (CC-BY 2.0 FR, attribution Tatoeba and contributors).
• literary_800 — Extracted from the UD_Vietnamese-VTB test split (CC-BY-SA-4.0).

Reproducing from scratch

The build script is in nom-vn:

python benchmarks/data/spell_correction_eval/build.py

Deterministic via fixed seeds (light=42, heavy=43); same input → same exact 2,098 pairs.

License posture

Combined dataset is CC-BY-SA-4.0 (the most restrictive of the constituent sources, applied for safety to derivative works). Per-split licenses noted above; if you only need a permissive subset, use business_55_* (CC0) or formal_72_* (public domain) which carry no share-alike obligation.

Citation

@misc{nom_vn_spell_correction_eval_2026,
  title={Vietnamese spell-correction evaluation grid},
  author={Nguyen, Viet-Anh and {Neural Research Lab}},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/nrl-ai/vn-spell-correction-eval}}
}