DiacBench (v2)

An 11-language benchmark for diacritics, tone-mark and Arabic vowel-mark restoration, built by Olaverse Lab to evaluate diacnet, diactag and any other diacritic restoration system on equal footing.

Each config contains 1,000 sentence pairs: a diacritic-stripped input and a correctly diacritized reference. The task is to reconstruct the reference from the input.

from datasets import load_dataset
ds = load_dataset("olaverse/diacbench", "yo", split="test")
# {"input": "se eranko naa si gbo o?", "reference": "ṣé ẹranko náà sì gbọ́ ọ?", ...}

v1 = load_dataset("olaverse/diacbench", "yo", split="test", revision="v1")   # original v1 references

What's new in v2

1. Corrected references for Yorùbá, Igbo and Hausa (and a handful in the other languages). A v1 audit found that part of every model's measured error on those three languages was error in the references themselves: missing underdots and wrong tones in Yorùbá, missing dotted vowels in Igbo, missing hooked letters in Hausa (e.g. kasar → ƙasar, kudi → kuɗi).
2. Arabic (ar): 1,000 fully diacritized Classical Arabic sentences, with a second reference for the no case endings setting.
3. New columns so nothing is hidden: reference_v1 (the original reference) and corrected_words (how many words changed in that sentence).

Config	Sentences corrected	Words corrected
yo	220	279
ha	121	166
ig	121	133
vi	15	16
es / pt / fr / tr / pl / it	6 / 3 / 2 / 2 / 1 / 1	15
ar (new)	–	–

Every change is listed in changelog.csv (config, row, word, v1 form, v2 form, reason). Only diacritics changed: for every corrected sentence, the stripped reference is identical to v1, so input is unchanged.

How the corrections were made

1. Second opinion. Two independent systems re-marked every stripped sentence: DeepSeek-V4-Pro (a large LLM) and diactag-2.0 (a character tagger).
2. Flags. A word was flagged when both systems produced the same form and it differed from the reference (1,537 flags; 6.9% of Yorùbá words, 2.4% Igbo, 1.1% Hausa, ≤0.6% elsewhere).
3. Independent judge. A third model (Claude Sonnet 5.5) saw each flagged word in its sentence and chose the reference form, the flagged form, both, or neither, with a confidence. Only high-confidence choices of the flagged form became corrections (609 of 1,537); everything else kept the v1 reference.
4. Manual check. A sample of verdicts was reviewed by hand. On 30 Yorùbá flags, all 12 accepted corrections matched the manual judgement; across ~40 accepted corrections checked in five languages, one was wrong and was reverted (Vietnamese như thể, "as if").

Caveat. Corrections were only proposed where the two re-marking systems agreed, so v2 cannot fix errors both of them share, and its gains favour systems that behave like them. Report v1 and v2 side by side when comparing systems. The corrections are not native-speaker verified.

Effect on one system (diactag-2.0), diacritic error rate (DER) / exact match:

	yo	ig	ha	vi	European (6)
v1	0.0787 / 0.076	0.0162 / 0.492	0.0044 / 0.741	0.0141 / 0.626	within 0.0004 / 0.007 of v2
v2	0.0729 / 0.107	0.0137 / 0.561	0.0025 / 0.833	0.0137 / 0.633

Languages

Code	Language	Diacritic type	Source	Licence
es	Spanish	accents (é, ñ, ü)	Tatoeba	CC-BY 2.0 FR
fr	French	accents (é, è, ç)	Tatoeba	CC-BY 2.0 FR
it	Italian	accents (à, è, ù)	Tatoeba	CC-BY 2.0 FR
pt	Portuguese	accents (ã, ç, õ)	Tatoeba	CC-BY 2.0 FR
pl	Polish	special letters (ł, ż, ą)	Tatoeba	CC-BY 2.0 FR
tr	Turkish	special letters (ı/İ, ş, ğ)	Tatoeba	CC-BY 2.0 FR
vi	Vietnamese	tone + vowel marks	Tatoeba	CC-BY 2.0 FR
ig	Igbo	dotted vowels (ị, ọ, ụ), ṅ	MasakhaNEWS	CC-BY 4.0
yo	Yorùbá	tone marks + underdots	MENYO-20k test split	CC-BY-NC 4.0
ha	Hausa	hooked consonants (ɓ, ɗ, ƙ)	MasakhaNEWS	CC-BY 4.0
ar	Arabic (Classical)	harakat, tanwīn, sukūn, shadda	Fadel et al. 2019 test split (HF copy)	MIT

Arabic config. Sentences of 30–200 characters, the first 1,000 of the Fadel test split. input strips harakat, tanwīn, shadda, sukūn and dagger alif but keeps hamza letters (أ إ آ ؤ ئ), which are spelling. reference is full diacritization including case endings; reference_nocase drops the vowel / tanwīn / sukūn of each word's last letter (shadda kept), for systems that restore everything except the grammatical endings. The Fadel dataset is derived from the Tashkeela corpus, which many Arabic diacritizers train on, so check for training overlap when comparing systems.

How it was built (unchanged from v1)

Diacritics are stripped deterministically (Unicode NFD + removal of combining marks, plus folding for letters with no combining form, e.g. ɓ→b, đ→d, ı→i; Arabic as above) to produce input from reference. Latin-script sentences were filtered to 30–200 characters, required to contain at least one diacritic, deduplicated and sampled with seed 42.

Evaluating a model

1. Feed each input to your model (for ar, score against reference or reference_nocase).
2. Report the diacritic error rate over letters that can carry a mark, plus WER and exact match.
3. Check that the output strips back to the input; generative systems sometimes change letters.
4. Report a copy-input baseline as the floor.

Known limitations

• Domain skew: Tatoeba is short and conversational, MasakhaNEWS is news, MENYO-20k is mixed translation data, Fadel is Classical Arabic (religious and legal texts). Compare systems within a language.
• Residual reference noise: v2 fixes errors that two systems agreed on; errors they share remain.
• Hausa is not tonal: its diacritics are hooked consonants, a different phenomenon from tone marks.
• No Modern Standard Arabic config yet; for MSA, see SadeedDiac-25 (WikiNews and new benchmark subsets).

Citation

@misc{diacbench,
  title  = {DiacBench: A Diacritics Restoration Benchmark},
  author = {Olaverse},
  year   = {2026},
  url    = {https://huggingface.co/datasets/olaverse/diacbench}
}

Licence

CC-BY-2.0 overall, except yo (CC-BY-NC-4.0, inherited from MENYO-20k: non-commercial use only) and ar (MIT, from Fadel et al.). Attribute Tatoeba, MasakhaNEWS, MENYO-20k and Fadel et al. as the underlying sources.