What LLM Support for Yoruba Actually Looks Like Today

Every large model will answer you in Yoruba. That is not the same as supporting it. The distance between the two is measurable, it is concentrated in one specific part of the orthography, and you can find out where your model sits in about five minutes.

What “support” has to mean

A vendor listing Yoruba on a language page is claiming almost nothing. Language support is not one property; it is at least five, and a model can pass three of them while failing the two that matter for your product:

• Encoding. Does the tokenizer represent Yoruba text without shredding it into bytes? This is the cheapest to satisfy and the one most often mistaken for the others.
• Comprehension. Can the model follow an instruction given in Yoruba, rather than only translating Yoruba into English internally and answering the English?
• Generation. Is the output grammatical, correctly toned, and idiomatic rather than calqued from English word order?
• Factuality. Does the model hold the same accuracy in Yoruba that it holds in English on the same question? It usually does not — see why low-resource languages hallucinate more.
• Safety. Do the refusal and moderation behaviours trained mostly on English transfer at all?

Yoruba today generally clears the first, partly clears the second, and is unreliable on the third, fourth and fifth. The reason is the same in every case, and it is a supply problem rather than anything about the language.

The numbers that are actually published

Ethnologue puts Yoruba at roughly 45–47 million speakers, mostly in south-western Nigeria, Benin and Togo, which places it among the most widely spoken languages in Africa. That is the demand side, and it is the number everyone quotes.

The supply side is harder, because no current frontier lab publishes a per-language breakdown of its training corpus. The last one that did was OpenAI, in the GPT-3 paper of May 2020, whose appendix listed the language distribution of its filtered Common Crawl and showed English at over ninety per cent of documents with more than a hundred other languages sharing the remainder (Brown et al., 2020). Nothing comparable has been published for a frontier model since. Anyone quoting you “Yoruba is 0.0X% of GPT-5’s training data” is making it up.

What you can check are the proxies the corpus is built from. The Common Crawl Foundation publishes per-language statistics for every crawl it runs (Common Crawl crawl statistics), and W3Techs runs a continuous survey of the content languages of the top ten million websites (W3Techs). On both, Yoruba sits far below one tenth of one per cent, while languages with a fraction of its speaker population sit above it. The Wikipedia editions tell the same story in a form you can count directly: compare the Yoruba edition against the English one on the list of Wikipedias. The ratio is roughly two orders of magnitude, in article count, for a language with tens of millions of speakers.

Where it breaks first: tone and underdots

Yoruba writes two things with diacritics that most Latin-script languages do not write at all. The acute and grave accents mark high and low tone, with mid tone unmarked; and a subscript dot distinguishes three vowel and consonant pairs — ẹ from e, ọ from o, ṣ from s. Both are load-bearing. The commonly cited demonstration is the set igbá (calabash), igba (two hundred), ìgbà (time) and ìgbá (garden egg): identical letters, different words, separated only by the marks.

Now consider what happens to that on the web. A large share of Yoruba text online is written without diacritics, because typing them requires a keyboard layout most people do not have installed. Undiacritised Yoruba is what gets crawled, so it is what dominates the training distribution, so it is what the model has learned to produce. The failure is therefore not random noise: models tend to emit fluent, confident, flat Yoruba, with the tone marks either absent entirely or applied inconsistently within a single paragraph. To a non-speaker the output looks fine. To a reader it is ambiguous at exactly the points where meaning is carried.

A second, subtler failure follows from the same cause: the model has seen far more Yoruba-in-English-syntax than Yoruba prose, because much of the crawled material is translation. Output often keeps English clause order and English discourse connectives, producing something a speaker recognises as translated rather than written. Neither problem is visible in a token-level metric, which is part of why they persist.

A five-minute check on your own model

Do not take a vendor’s language list as an answer. The following is cheap, needs no benchmark harness, and tells you which of the five properties above your candidate model actually has:

1. Ask for a paragraph of original Yoruba prose on a concrete topic. Count the tone marks. Zero marks across a full paragraph means the model is producing undiacritised Yoruba and cannot be trusted for anything published.
2. Give it a diacritised sentence and ask it to explain the meaning in Yoruba. If the answer switches to English unprompted, comprehension and generation are not both present.
3. Ask the same factual question twice, once in English and once in Yoruba, and compare the answers. Divergence is the factuality gap, not a translation artefact.
4. Ask it to restore the diacritics on an undiacritised sentence you supply. This is the single most useful Yoruba task a model can do well, and the one that most clearly separates models.

What Yoruba is usable for right now

The honest position is that Yoruba is at the stage where a model is a usable assistant to a Yoruba speaker and an unsafe substitute for one. Classification, routing, sentiment and rough gisting of user-generated Yoruba are workable, because they tolerate the diacritic loss. Anything user-facing and published — a generated notification, a translated help article, a synthesised answer — needs a speaker in the loop, and the loop is cheaper if you build the review step in from the start rather than after the first complaint. The participatory-translation work published by the Masakhane community (Nekoto et al., Findings of EMNLP 2020) is the reference design for that loop, and it is more useful reading than any model card.

One further thing worth deciding deliberately: Yoruba is a dialect continuum, and the standard written form rests on the Ọyọ variety. A model trained on what is online has learned standard Yoruba, so output aimed at speakers elsewhere reads as correct but not theirs. If your users are concentrated in one region, that is a product decision to make up front rather than one to discover from complaints.