Pilots
How a pilot works
A fixed-scope pilot on your own Arabic data, with gold-standard QA and a quality report, usually within two weeks. Then you decide whether to scale.
A pilot is the fastest honest way to judge an Arabic data vendor: a fixed scope, on your own data, measured against a gold standard. It usually runs within two weeks, and you decide whether to scale based on the quality report, not on a sales deck.
What a pilot includes
- A fixed scope. One task, agreed dialects and domain, a set volume and a clear definition of done.
- Your own data. Samples from your real distribution, not a demo set, handled under the defaults on our security and data handling page.
- Gold-standard QA. Work is calibrated against gold standards, with adjudication and rework loops and an audit trail on every task.
- A benchmark and quality report. What was delivered, how it scored against the gold standard, and what we'd change at scale.
How it runs: Vet, Train, Produce, QA, Deliver
- Vet. We match contributors who have passed dialect and domain screening for your task: native speakers, linguists or licensed professionals where the domain needs them.
- Train. Contributors calibrate on your guidelines and gold-standard examples before they touch production items.
- Produce. Annotation, human feedback, evaluation or red-teaming on the agreed sample.
- QA. Adjudication against the gold standard, with rework where items fall short.
- Deliver. Clean, structured data in the agreed format, plus the quality report.
What we need from you
| The task | What the data is for, and the labels, ratings or outputs you need. If you don't have guidelines yet, start from our Arabic annotation guidelines template. |
| Dialects and varieties | Which varieties your users speak or write (for example Najdi, Hijazi, Gulf, Egyptian, Levantine, MSA). “Arabic” alone is not enough to scope quality; see why dialect coverage matters. |
| Domain | Banking, healthcare, legal, government, telecom, media or retail, and whether licensed experts are needed. |
| Modality and volume | Text, audio, image or video, and a rough pilot and production volume. |
| Data constraints | Residency, on-prem or private-cloud requirements, and anything your security review needs. |
| Success criteria | What result would make you scale: the quality bar, the review process and who signs off. |
Pilots we run
- Arabic data annotation and labeling: NER and spans, intent and sentiment, classification.
- Arabic RLHF and preference data, and Arabic LLM training data.
- Arabic LLM and agent evaluation and Arabic red teaming.
- Arabic speech data and transcription, Arabic OCR annotation and dialect data.
After the pilot
If the quality report meets your bar, we scale in the way that suits your team: managed, embedded or enterprise. If it doesn't, the report shows exactly where Arabic quality breaks for your task.
Ready to scope one? Use the pilot scoping form. It takes about two minutes and opens an email to hello@bayanatlabs.com with your answers.