Instruction examples
Requests, context, and reviewed answers that include follow-up questions, changing instructions, and the right register.
Bayanat Align
Arabic data for better assistant responses.
Bayanat Labs (Riyadh) supplies Arabic instruction, preference and response-repair data: license existing sets or commission examples from a team of 50,000+ expert annotators across 25+ Arabic varieties.
اقرأ بالعربيةArabic alignment sets are also available to license with our smartphone speech dataset. Ask us about the available versions and counts.
Is Bayanat Align the right fit?
A good fit if
- You are tuning an Arabic assistant and need instruction, preference or response-repair records.
- You want judgments that keep correctness, naturalness, grounding, dialect and safety separate.
- You need examples built around your own policies, terminology and register.
What you can check before you license
We separate what is confirmed on this page, what we send on request, and what is agreed for each project.
| Item | Status | Details |
|---|---|---|
| Licensable alignment sets | Confirmed | Existing Arabic alignment sets can be licensed alongside our Arabic smartphone speech dataset. Versions and counts on request. |
| Record types | Confirmed | Instruction examples, preference comparisons and response repairs are separate modules. |
| Ties and reviewer disagreement | Confirmed | Kept visible in preference records instead of forced into one winner. |
| Open worked example | Confirmed | Our free, synthetic Arabic Preference Calibration Kit shows the record format: a rubric with safety and accuracy gates, nine gold pairs across MSA, Gulf, Egyptian and Levantine Arabic, a tie, a no-gold disagreement case and a both-unacceptable case, and a DPO-ready export. See the calibration kit |
| Sample records, rubric and reviewer decisions | On request | Review them before choosing the final delivery. |
| Dialect and register mix | On request | Per module, with the specifications. |
| Policy-based custom examples | Per project | Written from documents you are authorized to share, in your register. |
| Held-out evaluation set | Per project | A separate set to compare model versions without testing on training examples. |
What we can build together
Preference comparisons
Separate judgments for correctness, naturalness, grounding, dialect, safety, and concision. Keep ties and reviewer disagreement visible.
Response repairs
Flawed answers paired with specific problems and corrected responses. Evidence-based tasks include a way to check the answer.
Dialect and register
Interactions across MSA, regional Arabic, and mixed Arabic-English terminology, with meaning preserved when the requested register changes.
How the license covers each use
Training rights do not automatically include display, retrieval or redistribution. Each use is set in the agreement.
Model training
Permitted use is set in the agreement for the modules you license.
Your source material
Custom examples use documents you are authorized to share. Contributors work under NDA with least-privilege access.
Evaluation separation
Evaluation items stay separate from training deliveries.
How we put the data together
Start with your assistant
Tell us which requests it handles and where it struggles. We help select existing data and identify missing examples.
Review a sample
Check the records, writing guidelines, and reviewer decisions before choosing the final delivery.
Check improvement
Add a separate evaluation set to compare model versions without testing on the training examples.
For example: When the customer changes their mind
A customer corrects an address, then gives two conflicting delivery times. The assistant needs to keep the corrected address and clarify the time.
Scope your request
Send us these details and we’ll send the available specifications, the licence wording, and a sample where one is available, so your team can check the fit before licensing. If you need annotation added, we scope a pilot with gold-standard QA and a quality report, usually within two weeks.
- What your assistant does and where it struggles
- Record type: instruction, preference or repair
- Dialect and register
- Volume needed
- Timeline
Common questions
How is preference data different from instruction data?
Instruction examples show a good response to a request. Preference records compare candidate responses and explain which is better on specific dimensions, such as correctness or natural Arabic. You can license or request these as separate modules.
Can the examples use our policies and terminology?
Yes. Custom work can use documents you are authorized to share, your preferred register, and the situations your assistant handles. We review sample records and the annotation guidelines with you before production.