Arabic datasets for
the AI you’re building.
Bayanat Labs (Riyadh) supplies Arabic AI data in nine product families: 30,000 hours of Arabic speech, 2,700 Arabic books, 400,000 STEM Q&A items and 200,000 MSA NER sentences to license, plus custom collection, annotation and evaluation drawing on a team of 50,000+ expert annotators, the largest in the Middle East, across 25+ Arabic varieties.
اقرأ بالعربيةSpeech & text
Start with a collection you can license, then add the coverage you need.
Bayanat Voice
30,000 hours of Arabic speech, including two separate 1,000-hour Saudi collections, for speech recognition, conversations and voice agents.
Datasets + custom projects Books, text & question-answer dataBayanat Knowledge
2,700 Arabic books, 400,000 STEM Q&A items, 200,000 MSA NER sentences, 180,000 Arabic regional Q&A pairs and edited MSA for your language model.
Datasets + custom projects Studio recordings & expressive speechBayanat Voice Studio
Saudi expressive speech by professional voice actors and a 2,000-prompt, 40-speaker studio set, with voice-use rights agreed for your project.
Datasets + custom projectsTraining & evaluation
Help your model give better answers, take useful actions, and pass the tests that matter to you.
Bayanat Align
Examples, preferences, and response repairs that help assistants answer naturally and follow instructions.
Datasets + custom projects Tasks, tool use & agent trainingBayanat Agents
Arabic tasks and tool-use demonstrations, with custom environments that check whether the job got done.
Datasets + custom projects Test how your Arabic AI performsBayanat Bench
Find where your Arabic model struggles, from dialect understanding to completing a customer’s request.
Datasets + custom projectsDocuments, search & vision
Have a collection built around the documents, images, and questions your product handles.
Bayanat Documents
Read invoices, forms, and contracts with labeled text, layout, and fields.
Built to order Search relevance & grounded answersBayanat Search
Train retrieval with Arabic queries, relevance judgments, and answers linked to the evidence.
Built to order Arabic understanding in images & videoBayanat Vision
Arabic scene text, visual questions, and video events collected around your application.
Built to orderWhat you can verify on this site
For each family: what is confirmed now, and what we send on request. We never show a sample that has not been cleared for publication.
| Family | Confirmed on this site | On request |
|---|---|---|
| Bayanat Voice | Hours per collection; Transcripts and delivery schema | Speakers, dialect mix and collection dates; Recording source; Audio format and sample rate |
| Bayanat Knowledge | Counts per collection; Edited MSA coverage and rights | Title, publisher and subject lists; Token counts; NER label taxonomy and span rules |
| Bayanat Voice Studio | Studio Prompt Speech counts; Saudi Expressive Speech; Open inspection kit (text only, not a sample) | Expressive collection size and speakers; Prompt-to-speaker mapping; Recording setup and audio format |
| Bayanat Align | Licensable alignment sets; Record types; Ties and reviewer disagreement | Sample records, rubric and reviewer decisions; Dialect and register mix |
| Bayanat Agents | Licensable trajectories; Research method | Sample trajectory |
| Bayanat Bench | Licensable evaluation sets; Held-out separation; Published research | Sample report |
| Bayanat Documents | Source handling | Example of the label format |
| Bayanat Search | Relevance scale; Hard negatives; Unanswerable questions | Sample judgments and reviewer reasoning |
| Bayanat Vision | Real and generated images; People in images | Example of the label format |
What to include in a data request
A clear request gets you to the right collection or a scoped pilot faster.
- The task: what the model must learn or be tested on, and the modality (speech, text, documents, images or video).
- Existing or custom: whether a licensed collection could work, or the data needs to be built around your product.
- Language coverage: Arabic varieties, register, mixed Arabic-English, and where your users are.
- Volume and format: hours, records or pages, file formats, and the schema your pipeline expects.
- The uses to license: training, evaluation, retrieval, display, redistribution or synthetic voice, each named separately.
- How you will decide: the sample or pilot result that would let you proceed, and your timeline.