olaverse/diacnet-1.1-corpus

DiacNet 1.1 Corpus

Training corpus for diacnet-1.1, a diacritic-restoration model covering 10 languages: Yorùbá (yor), Igbo (ibo), Hausa (hau), Vietnamese (vie), Polish (pol), Turkish (tur), Portuguese (por), Spanish (spa), French (fra), and Italian (ita).

This is a cleaned, density-filtered, deduplicated sentence corpus sourced entirely from commercially-licensed data — built to fix known issues in diacnet-1.0: dropped/incomplete diacritics on some languages (a Unicode normalization + truncation bug) and unresolved tonal/diacritic ambiguity in Yorùbá and related languages.

The DiacNet 1.1 data family

Dataset	Role
diacnet-1.1-corpus (this repo)	Clean diacritized sentences — the raw material for training
diacnet-1.1-gloss	Ambiguity lexicon & English gloss tables derived from this corpus — powers the gloss-conditioning feature
diacnet-bench (planned)	Independent, human-verified evaluation benchmark with a hard-ambiguity subset

Sources & licensing

Source	License	Role
FineWeb-2	ODC-By 1.0	Primary web-text source, all 10 languages
Wikipedia (20231101 dumps)	CC BY-SA 4.0	Secondary, higher-trust source, all 10 languages

Both sources permit commercial use with attribution. No non-commercial-only sources (e.g. MENYO-20k, JW300) were used. This dataset is released under ODC-By 1.0, the more restrictive of the two upstream licenses; downstream users must attribute FineWeb-2 and Wikipedia per their respective terms.

Build pipeline

1. Stream & sentence-split — FineWeb-2 and Wikipedia streamed per language; sentences split with wtpsplit (sat-3l-sm), a neural multilingual segmenter (regex/pysbd splitters were tried first and rejected — pysbd doesn't support vie/tur/por, and regex mis-split abbreviations).
2. Normalize — every sentence normalized to Unicode NFC. This directly addresses a v1.0 bug where mixed NFC/NFD training targets likely caused inconsistent diacritic output in Vietnamese.
3. Length filter — dropped below 12 characters or above 512 UTF-8 bytes (covers Vietnamese NFD/NFC byte-length inflation that caused truncation in v1.0's 256-byte limit).
4. Diacritic-density filter — the key quality gate. For each language, computed the fraction of diacritic-eligible characters that actually carry marks, per sentence, then calibrated a per-language minimum threshold from the bimodal histogram of (undiacritized-web-text spike) vs. (properly-marked-text mode). Undiacritized/ASCII-ified sentences — common in Yorùbá, Igbo, and Hausa web text — are the single biggest quality risk for this task and are explicitly filtered out rather than trained on.
5. Dedup — exact dedup, plus stripped-form dedup (near-duplicate sentences that differ only in diacritic marks are collapsed, keeping the highest-density variant). This also guarantees no overlap between train/validation and the test split.
6. Split & sample — 2,000 sentences/language reserved for validation and test (disjoint, held out before any training use). High-resource languages (vie, pol, tur, por, spa, fra, ita) capped at 400k–500k sentences for training; low-resource languages (yor, ibo, hau) keep everything that passes the filters.

Dataset structure

from datasets import load_dataset
ds = load_dataset("olaverse/diacnet-1.1-corpus")

DatasetDict({
    train:      Dataset({ features: [...], num_rows: 7,020,064 })
    validation: Dataset({ features: [...], num_rows: 20,000 })
    test:       Dataset({ features: [...], num_rows: 20,000 })
})

Columns

Column	Type	Description
text	string	Fully diacritized sentence (NFC-normalized), used as the training target. The stripped input (what the model receives at training/inference) is derived from this at training-example-build time, not stored here.
lang	string	3-letter language code: yor, ibo, hau, vie, pol, tur, por, spa, fra, ita
source	string	fineweb2 or wikipedia
density	float	Diacritic density score (fraction of eligible characters carrying a mark) used for the quality filter
bytes	int	UTF-8 byte length of text

Language breakdown (train split)

lang	sentences	median bytes	% from Wikipedia	mean density
yor	1,897,164	119	3.6	0.223
ibo	1,850,691	134	16.0	0.219
hau	400,209	123	31.7	0.046
vie	496,000	130	0.0	0.631
pol	396,000	91	0.0	0.137
tur	396,000	95	0.0	0.368
por	396,000	109	0.0	0.081
spa	396,000	129	0.0	0.054
fra	396,000	121	0.0	0.084
ita	396,000	132	0.0	0.039

Mean density is not comparable across languages — it reflects how many characters can carry diacritics in that language's orthography, not corpus quality (e.g. Hausa naturally has few diacritic-eligible characters per sentence; low density there is expected and correct, not a sign of undiacritized text).

A cross-language analysis of the ambiguity in this corpus (which stripped forms map to multiple diacritized words, and of what kind) is published in diacnet-1.1-gloss; headline finding: Vietnamese ambiguity is near-total and lexical, Spanish/Polish/French ambiguity is mostly grammatical, Italian has essentially none, and Yorùbá/Igbo/Hausa corpus variance is dominated by tone-marking completeness with a small high-value core of true tonal minimal pairs.

Known limitations

• Hausa is the smallest split (~400k vs. ~400k–1.9M elsewhere) and was the least-verified language in diacnet-1.0. FineWeb-2's Hausa subset was exhausted by the density filter before hitting the sampling cap; supplementary Hausa sources may be added in a future revision if diacnet-1.1 evaluation shows it's still the weak point.
• This corpus is not the evaluation benchmark. A separate, human-verified benchmark (diacnet-bench) sourced independently (including a "hard" subset of known-ambiguous sentences) is planned for measuring model quality — this corpus and that benchmark are disjoint by design (see test split / dedup above), but diacnet-bench itself will draw on different, more strictly redistribution-safe sources.
• Sentences are web/encyclopedia text; the corpus does not itself include gloss (word-sense) annotations — those live in the companion diacnet-1.1-gloss dataset and are injected in a downstream training-example-construction step.
• Diacritic-density filtering is a heuristic proxy for "correctly diacritized," not a guarantee — some noise likely remains, particularly near each language's calibrated threshold.

Intended use

Training data for sequence-to-sequence diacritic restoration models (e.g. byte-level T5 variants). Not intended as a general-purpose multilingual text corpus — sentences were selected specifically for diacritic density and length, which biases against very short/simple sentences in some languages.

Citation

If you use this dataset, please cite the upstream sources:

FineWeb-2: Penedo et al., HuggingFaceFW/fineweb-2
Wikipedia: Wikimedia Foundation, wikimedia/wikipedia dumps (20231101)

and this dataset:

@dataset{diacnet_1_1_corpus,
  author = {olaverse},
  title = {DiacNet 1.1 Corpus},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/olaverse/diacnet-1.1-corpus}
}