AletheiaAING/Data-Quality

African-Language Data Quality

Seven datasets built around one question: how do you tell good African-language text from bad, when there is no reference translation to compare against? They fall into three families, which do not share a schema — load them by config name.

config	rows	languages	what it is
default	493,831	71	human text, corrupted by a known function by a known amount
270m_inlang_full	57,465	71	model-generated text, in-language prompts
270m_inlang_b	45,560	71	the Box B slice of the above, on its own
270m_in_lang	1,641	68	same regime, separate box
270m_pool_b	1,247	20	English-instruction prompts — ~99% failed, kept as a floor class
270m	3,742	20	earliest generation run, English instruction prompts
diacritics_sft_v2	163,480	40	diacritic restoration SFT, + AfriADR train
diacritics_sft	135,116	39	the same, before AfriADR was added

from datasets import load_dataset
d = load_dataset("AletheiaAING/Data-Quality", "270m_inlang_full")

---

Family 1 — graded degradation (default)

493,831 documents across 71 languages, each either clean or corrupted in exactly one named way, by a known amount: 7 dimensions (vocabulary, grammar, format, discourse, cultural_load, code_switch, idiomatic) x 5 severities (none, subtle, light, moderate, heavy). Because the corruption is a function, every row carries an exact label rather than an opinion.

Why degradation and not provenance. Provenance was tried first and failed. Over 120+120 documents per language, curated sources separated from web crawl in the wrong direction on the graded dimensions — vocabulary AUC 0.000 for Swahili, complexity 0.013 — because African Wikipedia is largely short templated stubs while crawl contains long-form prose. Those metrics were reading genre, not quality.

Family 2 — real model output (270m*)

The degradation family is corrupted by code; these rows are corrupted by a model. Keeping them apart is what makes the important question answerable: does a quality detector work on real generation, or has it only learned to recognise my corruption functions? Train on default, evaluate here.

270m_inlang_full is the one to use. Each prompt is a short (12–35 word) sentence of real corpus text in the target language, which a base google/gemma-3-270m continues. A base checkpoint is deliberate: with no chat template it continues text natively, where an instruction-tuned one answers conversationally in English — english-ness 0.50 against 0.10 for this setup, with real corpus text at 0.063. Continuations restating their own prompt (>=50% 5-gram overlap) are dropped, so a copy of real corpus text is never stored as machine output.

270m_inlang_b is the Box B portion of 270m_inlang_full, published separately so the two machines' output can be used or audited independently. The 270m run was sharded across a 9-slice fleet on CPU (24–42 tokens/second, ~12 hours per worker), and the shard prefixes do not record which slice each box took. Measured rather than assumed: Box A contributed 11,905 rows and Box B 45,560, with zero shared prompts — the two slices were genuinely disjoint, and 270m_inlang_full contains no duplication.

270m_pool_b is published as a negative result. English instruction prompts against a base checkpoint produced ~99% English boilerplate carrying an African language code. Wrong-language output is a real floor class for a quality estimator, but never pool it with in-language rows or the model learns that Yoruba looks like English.

Rows where tier1_is_text is false carry no tier2_* scores — the gate short-circuits grading. They are still informative: they failed on digit runs and repetition.

Family 3 — diacritic restoration (diacritics_sft*)

Supervised pairs for restoring diacritics, prompt format:

Diacritize this text: {stripped} <tgt> {diacritized}<eos>

diacritics_sft_v2 is current: 163,480 rows, 40 languages, including the train split of masakhane/AfriADR (apache-2.0). Its dev and test splits are deliberately excluded so they remain a usable external benchmark; overlap with its test split is at most 0.5%, in Igbo.

v2 carries a strip_style column, because two different operations were being conflated:

• marks_only — removes combining marks: tone (à á â ǎ) and the Yoruba/Igbo under-dot (ọ ẹ ṣ). This is the classic task, and AfriADR's definition.
• ascii_fold — additionally substitutes extended Latin letters (ɓ ɗ ƙ ƴ ɛ ɔ ŋ ʋ ɖ ƒ ə) for ASCII, modelling text typed without them. MasakhaNER 2.0's Hausa is fully ASCII-ised: 0.00 hooked consonants against 34.65 per 1k words in our corpus.

Which is correct is per language. Yoruba and Igbo mark with combining characters, so both coincide. Fon and Ghomálá' write ɔ ɛ ɖ ə ŋ as distinct letters of the alphabet — folding those destroys a letter rather than removing a diacritic. Against AfriADR's own stripped side, marks-only agrees 398/400 on Ghomálá' where ascii-fold agrees 0/400.

This is not cosmetic: a model trained only on ascii_fold scores at or below a copy-the-input baseline on Fon and Ghomálá', while reaching 0.35 exact match on Igbo, where the two definitions agree. See the model card.

Identity rows are isiZulu, isiXhosa and Swahili — orthographies carrying no diacritics, at 0.00005, 0.00013 and 0.00009 sites per character. The correct target IS the input. Without them, a model trained only on diacritic-rich languages over-applies marks to languages that take none.

---

Provenance and licence

The 270m* and diacritics_sft* families derive from this project's own pretraining corpus and, for v2, from AfriADR (apache-2.0).

The default config is built from mixed upstream sources, and the YAML licence declaration above is inherited from this repository's first publication rather than verified per source. If you intend commercial use, confirm the upstream licences for that config yourself — the graded degradations are ours, the underlying documents are not necessarily.

Generated with Claude Code.