Bayanat Labs press kit

About Bayanat Labs

Short boilerplate. Bayanat Labs is a Riyadh-based Arabic AI data company. It licenses ready-made Arabic speech and text datasets, and sources, annotates, aligns and evaluates Arabic data across 25+ dialects: text, speech, image and video. Every label Bayanat produces comes from vetted native speakers and domain experts, is measured against gold standards, and is hosted in-region by default.

Long boilerplate. Bayanat Labs builds the human data that makes Arabic AI work. The company does data only and is model-agnostic: dialect-aware annotation, human feedback and RLHF, domain-expert data (medical, legal, financial), and independent Arabic LLM and agent evaluation. The work is done by vetted native speakers and licensed professionals across 25+ Arabic varieties. Contributors pass dialect and domain screening, work is calibrated against gold standards, and every task is adjudicated and audited. Data is hosted in-region by default, with on-prem and private-cloud options. Its catalog has nine product families: Voice, Voice Studio and Knowledge offer ready-to-license collections, including two separate 1,000-hour Saudi speech collections, 2,700 Arabic books and 400,000 STEM questions; Align, Agents and Bench add alignment, agent and evaluation data; Documents, Search and Vision are built to order. Its open research includes LabelBench, an audit of whether AI can label Arabic data.

Key facts

LabelBench in four numbers

LabelBench asks when Arabic labeling can safely move from people to models. We ran one fixed zero-shot rubric on identical held-out samples and compared the models with small supervised baselines.

• 78% → 25.3%. A local language model (Qwen 2.5 3B) scored 78% on binary Arabic toxicity. On the real seven-label taxonomy, exact-match accuracy fell to 25.3% on the fixed held-out manifest.
• 22.4%, where chance is 20%. The same model reached 22.4% when asked to identify five Arabic dialect regions.
• 75% of Gulf examples labeled Egyptian. Its errors were not random: it labeled 70–80% of each non-Egyptian region's examples as Egyptian.
• Specialization and native data help. Arabic-specialized Fanar 1 9B did better on all three tasks (47.8% on dialect region), and a character n-gram model trained on 100 examples per region reached 69.8%.

The paper, results (CSV and JSON) and checksums are public, and the code is on GitHub and on PyPI; see how to cite or embed LabelBench.

Charts

• The annotation complexity cliff (PNG)
• Dialect confusion matrix (PNG)

Free to use with the credit “Source: Bayanat Labs, LabelBench v0.1” and a link to bayanatlabs.com/research/labelbench.

Story angles

• Can AI replace Arabic data annotators? Why a model that looks competent on yes/no labels collapses on real annotation work.
• The dialect blind spot. Why a model can mistake Gulf, Levantine, Iraqi and Maghrebi Arabic for Egyptian, and what that means for Arabic chatbots and moderation.
• Sovereign AI needs sovereign data. Why in-region, native-speaker data and evaluation matter for Gulf AI programs. Background: PDPL and Arabic AI training data.
• The cost of Arabic in AI. The tokenization tax: why the same content can cost more to process in Arabic.

Speaking topics

For podcasts, panels and conference programmes. Talks can be given in English or Arabic.

• Arabic annotation is not binary. What LabelBench shows about AI-assisted labeling, and why Safe Automation Coverage, not average accuracy, should decide what leaves the human queue.
• A fluent answer is not a completed transaction. Evaluating Arabic agents by the state they leave behind, not by their prose: the method behind the Arabic Agent Reliability Lab (methods beta).
• Who tests the Arabic that Gulf users actually speak? How dialect-specific red-teaming and evaluation are designed and staffed. Background: Arabic red teaming.
• Beyond the leaderboard. How a bank, hospital or telco can compare Arabic models on its own traffic with a small, dialect-stratified, human-judged evaluation. Background: evaluating Arabic LLMs.
• Noisy, code-switched and on the phone. What voice-agent teams should measure per dialect and per audio condition. Background: Arabic speech data.
• From توسيم to توصيف: the language of Arabic AI procurement. Writing RFPs, contracts and job posts that are correct and findable, using the official SDAIA glossary terms. Background: دليل المصطلحات.

Logos

• Logo mark, PNG, 512×512
• Logo mark, SVG (for light backgrounds)
• Logo mark, SVG (for dark backgrounds)

Please don't recolour or stretch the mark. The wordmark is set as “Bayanat Labs”.

Contact

Interviews, data questions and corrections: hello@bayanatlabs.com. For partnerships or a pilot, use the pilot scoping form, and see how a pilot works and how we handle data.