Arabic AI data for RLHF, DPO, and evaluation

What is Diraya

Diraya (Arabic: دراية) is a UAE-based supplier of Formal Modern Standard Arabic preference and evaluation data for post-training workflows — RLHF, DPO, and related preference learning — plus diagnostics that show where models fail.

We are not a Scale-scale marketplace. We do not claim thousands of anonymous crowd labelers. Inventory today is honest and small: 20 expert-audited preference pairs, a diagnostic on request, and larger packs by engagement. That is the offer. Everything else is roadmap.

Formal Modern Standard Arabic preference data

Our beachhead is Formal Modern Standard Arabic — the formal written Arabic used in schools, media, and official settings (sometimes abbreviated MSA; we spell it out because buyers mean different things by “Arabic”). Preference content is Arabic. Packaging and sales docs can be English or Arabic.

The failure mode we target: models that sound fluent but pick the wrong answer on grammar (nahw), morphology (sarf), diacritics (tashkeel), or syntax — case marks that rewrite who did what, morphology that flips agent and patient, vowel marks that turn praise into denial.

Dialects are Coming next, not sold as live inventory.

Why labs need Arabic RLHF and DPO data

Generic multilingual scrapes and lightly moderated chat data do not catch curriculum-grade Formal Modern Standard Arabic errors. Preference learning needs chosen vs rejected answers with a reason the preference is trainable — not just “B is better.”

Labs building Arabic AI systems — foundation model teams, evaluation stacks, Gulf and MENA localization programs — need expert-verified Arabic datasets where the gold answer is defensible against school-standard and official-register expectations. Arabic RLHF and Arabic DPO packs are one way to encode that signal. Model evaluation in Arabic needs the same rigor: pairs and diagnostics that expose silent meaning failures, not only fluency scores.

What we deliver

Core product language for Diraya today:

• Preference pairs — prompt · rejected (A) · chosen (B) · expert rationale, delivered as JSON or CSV for training stacks.
• Expert rationales — why the gold answer wins; usable in preference learning and as structured traces for supervised fine-tuning workflows (not a standalone multi-step chain-of-thought catalog).
• Evaluation support — preference rows that double as hard cases for Arabic model evaluation.
• Diagnostics — a short Formal Modern Standard Arabic fail sample on request, plus the option to request the full JSON/CSV pack.
• Reasoning traces — expert explanations attached to pairs; same honesty rule: not marketed as a separate product line beyond what packs include.

See a live sample row on the homepage sample, or download DIRAYA-AR-001-sample.json.

How expert review works

Lead reviewer for the Formal Modern Standard Arabic beachhead: a UAE MoE-accredited Arabic educator with 30+ years of classroom experience. We are expanding a faculty / academic review network. We do not market reviewers as PhDs unless they hold that title.

That credential is the right fit for school-standard and curriculum Formal Modern Standard Arabic. It is not sufficient to call Islamic Fiqh / Fatwa or personal-status rows enterprise-ready. Those domains need Sharia co-audit (or an explicit madhhab and jurisdiction disclaimer) before any buyer pitch — and they remain Coming next.

Beachhead vs Coming next

Live / core offer

• Formal Modern Standard Arabic preference pairs (grammar, morphology, diacritics, syntax)
• 20 expert-audited pairs in the current honest inventory
• Diagnostic on request · larger packs by engagement

Coming next (not sold as live)

• Islamic Fiqh & Fatwa preference (needs Sharia co-auditor before enterprise claims)
• GCC legal / personal status
• Dialects
• STEM / academic science preference
• Broader faculty / expert network

How to request a diagnostic

A diagnostic is a short Formal Modern Standard Arabic fail sample plus the option to request the full JSON/CSV pack. Use it to see failure modes before committing to a larger engagement.

Request diagnostic Back to Diraya home

Contact: set in config.js as DIRAYA_CONTACT — currently wired to the site inbox when configured.