Insights · 5 min read
Arabic-native training data: why translated data fails Gulf business
A translated task teaches a model how English sounds in another language.
Most multilingual training data starts in English and gets translated. It’s the cheapest way to cover many languages. It is also the fastest way to teach a model the wrong habits.
What translation loses
- Register: how formal a Gulf operations manager expects a reply to be is not how a casual English email reads.
- Phrasing: a colleague in Riyadh or Cairo doesn’t phrase a request the way an English sentence translated word for word does. Dialect, honorifics and Arabizi in chat all carry meaning.
- Mixing: real Gulf work switches between Arabic and English in the same thread. Translated data never does.
- Context: invoices, tax rules, honorifics, dates, number formats and approval customs differ by market.
- Reasoning: tasks built around English assumptions test translation, not understanding.
What native means for us
Our Guild writes every lesson first in the language it will be used in, by people who work in that language every day. We cover Arabic, formal and dialect, and the English, French and German the region trades in. When the same task is issued in several voices, each version is written from scratch.
Graded in every language
Native data still needs objective grading. For language tasks we combine exact-match checks on extracted facts with a weighted rubric that scores register, accuracy and completeness. Our samples page shows a fully worked Arabic example.
Businesses across the Middle East and North Africa will hand work to agents. Those agents should meet their colleagues in their own language, on their own terms.
See the format for yourself.