Bayanat Labs · بيانات لابز

Bayanat Labs is a Riyadh-based Arabic data company for AI. We collect, annotate, align and evaluate data across 25+ Arabic varieties, with vetted native speakers, linguists and licensed domain professionals. We are model-agnostic and work on data.

Bayanat Labs is not affiliated with bayanat.ae (the UAE open-data portal), the former Bayanat AI (now Space42), or the ARBML bayanat library.

Website · GitHub · LinkedIn

Open research and resources

• Arabic Reconciliation Bench — Original synthetic Arabic document reasoning: 5,600 generated families, 67,200 controlled views and 95,200 matched comparisons, plus 1,120 Arabic image pages. Includes exact fictional policies, independent software audits and reproducible evaluation; no human accountant validation. Data: CC BY 4.0. Inspect the evidence · Code.
• Arabic Entity Atlas — 2,296,957 Arabic-labeled Wikidata entities with source aliases, explicit ambiguity and component-disjoint retrieval partitions. CC0 data with reproducible code and audits. Explore the interactive sample.
• Arabic Retrieval Starter Kit — A fully synthetic Arabic/English retrieval and RAG evaluation fixture with 12 passages, 12 queries and 30 graded judgments. Includes Gulf and MSA questions, code-switching, superseded-policy hard negatives and unanswerable queries. CC BY 4.0. A pipeline fixture, not a representative benchmark or commercial-data sample.
• LabelBench — A reproducible audit of automated Arabic labeling, with an open evaluation harness, technical report and aggregate results. Code: Apache-2.0; Bayanat-authored results and paper source: CC BY 4.0. Third-party source data retains its own license and is downloaded from its original source.
• Arabic Agent Reliability Lab — A local harness that grades tool actions and resulting state, with 21 open synthetic commerce workflows and sanitized evidence receipts. Methods beta; colloquial examples are unvalidated pending independent native review. Code: Apache-2.0; open scenarios and published receipts: CC BY 4.0. Results describe disclosed whole-system configurations.

Work with us

Explore Arabic speech and text datasets, request custom collection and evaluation, or run a pilot. Commercial collection samples and datasheets are available on request.

Contact: hello@bayanatlabs.com