Rootspeak builds native-authored training and evaluation data in Hausa, Yoruba, and Igbo — three of West Africa's most-spoken languages, and three of the most under-resourced in AI training data today. This page is the full, current catalog: exactly what we produce, and exactly how each unit earns its way into a dataset we're willing to sell.
Two formats, built to the same standard, in all three languages. Both are live, tested pipelines — not a roadmap.
A native speaker writes a natural, fluent response to an assigned prompt — everyday situations, opinions, instructions, comparisons, and dozens of other categories, deliberately spread thin so no single topic starts to feel manufactured. Every response is original writing, never translated from English and never machine-generated.
A native speaker translates an English source sentence naturally into their language — not word-for-word, but the way a fluent bilingual speaker would actually say it. Back-translated and compared against the original source as part of the same quality process every unit goes through.
This is the actual pipeline every unit of work passes through — the same one our own operations dashboard is built around internally. Nothing here is aspirational; it's what genuinely has to happen before a unit becomes sellable.
Hausa, Yoruba, and Igbo were chosen for the same reason: tens of millions of native speakers each, real demand from AI labs building Nigerian and West African language coverage, and a fraction of the training data that population deserves relative to its size. We'd rather cover three languages this thoroughly than spread the same effort across a dozen we couldn't stand behind.
Contributors are paid per accepted unit, above local market rates, under signed agreements — never a race to the lowest bid.
Tell us what you're working on and we'll walk you through a sample matched to your use case — instruction data, translation pairs, or both.