Inside a Sovereign Arabic LLM Data Pipeline (Vision 2030 Scale)

Why Sovereign Arabic LLM Programmes Exist

Saudi Arabia's National AI Strategy, operating under SDAIA and aligned with Vision 2030, identified Arabic language AI as a strategic national capability. The assessment behind that designation is straightforward: Arabic is spoken natively by approximately 470 million people, but the global AI industry has invested less than 2% of its LLM training compute in Arabic language data. The result is a generation of commercially dominant AI systems — GPT-4, Claude, Gemini — that perform substantially worse in Arabic than in English, French, German, or Chinese.

The McKinsey Global Institute (2024) estimated that Arabic-language AI adoption could contribute USD $320 billion to MENA regional GDP by 2030. That value is unrealisable without LLMs that perform at parity with English systems on Arabic tasks. Sovereign programmes — SDAIA's Allam, TII's Falcon, Aramco's AceGPT — represent an institutional response: rather than waiting for foreign commercial providers to improve their Arabic capability, Gulf states are developing their own.

The strategic rationale also includes data sovereignty and PDPL compliance: training foreign commercial LLMs on Saudi government data, citizen interaction logs, or sensitive enterprise data requires exporting that data to foreign infrastructure. A domestically trained sovereign model processes and retains data within KSA, satisfying PDPL data residency requirements and reducing geopolitical dependency on foreign AI infrastructure.

The Five Stages of a Sovereign Arabic LLM Data Pipeline

Sovereign Arabic LLM data pipelines are not monolithic. They operate in five functionally distinct stages, each with different quality requirements and annotation expertise:

Stage 1: Pre-training data curation

Arabic Common Crawl provides the largest available source of Arabic web text — approximately 280 billion tokens in the CC-100 filtered corpus. However, unfiltered Arabic web data contains substantial proportions of low-quality content: scraped social media without context, duplicate news reprints, spam, and non-native Arabic written by non-Arabs (particularly in formal government announcements translated from English). Sovereign programmes apply annotation-assisted filtering:

• Dialect identification: each document classified by dialect (Khaleeji, Egyptian, Levantine, Maghrebi, MSA, unknown)
• Quality scoring: native Arabic annotators rate document quality on fluency, naturalness, and informativeness dimensions, producing labelled training data for a quality classifier
• Domain tagging: documents tagged by domain (news, government, literature, academic, social, commercial) to enable domain-balanced training batches
• Deduplication: exact and near-duplicate detection with human review of edge cases where semantic duplication is non-obvious

Current sovereign Arabic pre-training programmes at Vision 2030 scale are targeting 500 billion to 2 trillion tokens of filtered Arabic text for their primary training runs. The KAUST Arabic Web Corpus (2025) — a 1.2 trillion token filtered corpus produced under a national AI infrastructure grant — is publicly referenced as a data source for several sovereign programmes. Its production involved approximately 18 months of annotation-assisted quality filtering.

Stage 2: Supervised fine-tuning (SFT) instruction pairs

SFT instruction data transforms a pre-trained base model into an instruction-following assistant. For sovereign Arabic LLMs, this requires instruction pairs written natively in Arabic — not translated from English — across the full range of dialects the model must handle.

The failure mode of translated SFT data is well-documented: models trained on English-to-Arabic translated instruction pairs learn to respond in formal MSA regardless of the dialect of the user prompt. A Saudi citizen who asks a government services chatbot a question in Najdi Arabic and receives a response in formal MSA has had a technically correct but socially jarring interaction. The model has failed its use case.

Sovereign SFT programmes address this with dialect-stratified annotation: instruction-response pairs are commissioned in each target dialect by native speakers of that dialect. For a KSA-targeted sovereign model, this typically means:

• 35–40% Najdi Arabic instruction pairs (central KSA, Riyadh)
• 20–25% Hejazi Arabic (Jeddah, western KSA)
• 15–20% Gulf/Khaleeji (Eastern Province, GCC interaction contexts)
• 15–20% Modern Standard Arabic (formal, government, broadcasting)
• 5–10% code-switching (Arabic-English mixed, reflecting professional and business contexts)

Total SFT corpus size for a production sovereign KSA model ranges from 80,000 to 500,000 instruction pairs, depending on model scale and task coverage. The Allam programme (SDAIA, 2024) publicly disclosed that its SFT dataset required "thousands of native Arabic annotators across Saudi Arabia" — consistent with a 200,000+ instruction pair programme running over 12–18 months.

Stage 3: RLHF preference annotation

Reinforcement learning from human feedback (RLHF) and its variants (DPO, IPO) use pairwise preference data — human annotators selecting which of two model responses is better — to align model behaviour with human values and preferences. For sovereign Arabic models, preference annotation must be performed by annotators from the target culture, not generalised Arabic speakers.

Gulf Arabic cultural alignment requirements differ from those encoded in RLHF datasets trained primarily on Western or Egyptian Arabic annotators. A response considered appropriately humble or deferential in Western social contexts may read as sycophantic to Saudi users. A direct refusal of a religiously sensitive request must use appropriate Islamic framing. Responses about gender interactions, public behaviour, and government institutions must reflect KSA social and legal context accurately.

Sovereign RLHF programmes address this by:

• Requiring annotators to be Saudi or GCC nationals for KSA-targeted preference data
• Providing cultural alignment training to annotators before preference tasks
• Including religious scholars or cultural consultants in the preference rubric design
• Tracking preference divergence between annotator cultural subgroups as a quality signal

Stage 4: Safety evaluation and red-teaming

Sovereign Arabic LLMs deployed in government services and public-facing applications require safety evaluation that reflects KSA-specific safety requirements. This is structurally different from Western AI safety evaluation. Content that is harmful in Western contexts (extremist content, disinformation) is harmful in KSA context too. But the KSA context adds additional categories: content that is religiously sensitive under Islamic law, content that references the Saudi royal family inappropriately, and content relating to regional political sensitivities.

Red-teaming annotation for sovereign Arabic models involves:

• Arabic adversarial prompt generation in all target dialects — testing whether safety guardrails hold under dialectal variation
• Cultural harm labelling: classifying model outputs as safe, culturally inappropriate, or actively harmful by KSA social and legal standards
• Religious sensitivity review: qualified Islamic scholars reviewing model responses to religiously sensitive prompts
• Political sensitivity annotation: review of model responses to prompts about regional politics, government, and Vision 2030 by cultural consultants

Stage 5: Arabic benchmark construction

Sovereign programmes require evaluation benchmarks that test their specific deployment use cases rather than relying entirely on public benchmarks. ArabicMMLU, AlGhafa, and OALL provide useful cross-programme baselines, but they do not test sovereign-specific capabilities: citizen services task completion, government document understanding, Saudi legal and regulatory knowledge, or KSA cultural knowledge.

Sovereign benchmark annotation involves:

• Domain expert annotators: Saudi legal professionals annotating legal QA pairs, government document specialists annotating document understanding tasks, medical professionals annotating health information tasks
• Dialect-stratified evaluation sets: separate benchmark splits for each target dialect to surface dialect-specific capability gaps
• Cultural knowledge evaluation: questions about Saudi history, geography, law, culture, and Vision 2030 policy that require KSA-specific knowledge

Case Study: Government Services LLM Annotation Programme (KSA)

In 2025–2026, a SDAIA-aligned government AI programme developing a citizen services language model for Saudi e-government platforms required a production-scale annotation pipeline. The model needed to handle citizen queries in Najdi, Hejazi, and Khaleeji Arabic across government domains including civil status, traffic, healthcare, and tax services.

Annotation scope:

• 180,000 SFT instruction pairs across 12 government service domains, in 4 Arabic varieties
• 45,000 RLHF preference pairs with Saudi national annotators only
• 8,000 red-teaming adversarial prompts with safety classification labels
• 2,400-item domain-specific benchmark (civil status, traffic, healthcare, tax) with domain expert annotation

Pipeline structure:

• Annotator pool: 340 Saudi national annotators recruited across regions (Riyadh for Najdi, Jeddah/Mecca for Hejazi, Dammam/Khobar for Khaleeji/Eastern Province)
• Data handling: all data processed on KSA-resident infrastructure; annotation platform deployed within SDAIA-compliant cloud environment
• QA protocol: 10% sampling rate with senior reviewer at 5-level Likert quality rating; annotators falling below 0.72 kappa on calibration task replaced
• Timeline: 14 months from annotator onboarding to final dataset delivery

Model performance on internal benchmark (pre- vs post-annotation programme):

• Task completion rate: 41% → 79% (baseline was Allam-7B without domain SFT)
• Dialect naturalness rating (Saudi citizen panel): 2.8 → 4.5 / 5.0
• Safety compliance rate (red-team adversarial prompts): 67% → 94%
• Escalation rate (queries routed to human agent): 58% → 23%

The escalation rate improvement had direct operational cost implications: each human-agent escalation costs approximately SAR 45 (AUD $18) in operator time, and the deployment volume of 120,000 monthly queries meant the annotation programme paid back its total cost within 4 months of the model going live.

PDPL Compliance Architecture for Sovereign Data Pipelines

PDPL compliance in a sovereign Arabic LLM data pipeline is not a single control — it is an architecture. The key compliance requirements that shape pipeline design are:

The Data Flywheel Problem in Sovereign Arabic AI

Global commercial AI systems — GPT-4, Claude, Gemini — benefit from a data flywheel: user interactions generate implicit feedback signals that continuously improve the model through online learning, RLHF updates, and interaction log analysis. Sovereign Arabic programmes face a structural challenge: in the early deployment phase, the user base is smaller, the interaction volume is lower, and the feedback flywheel spins more slowly.

The operational response is to front-load annotation investment. Rather than relying on post-deployment user signals to improve dialect capability, sovereign programmes invest heavily in pre-deployment annotation — producing high-quality SFT and RLHF data before the model goes live, so it performs well enough on deployment day to generate genuine user engagement rather than user abandonment.

The Allam 7B model (SDAIA, 2024) demonstrated this approach: SDAIA publicly reported investing in over 60 billion tokens of Arabic-specific annotation and curation before releasing Allam, and committed to ongoing annotation programmes tied to government service deployment. The ArabicMMLU leaderboard showed Allam outperforming GPT-3.5 on multiple Arabic reasoning benchmarks at its launch — a result that would not have been possible without the pre-deployment annotation investment.

For Arabic AI teams outside sovereign programmes, the implication is that annotation investment early in the model lifecycle has a higher return than annotation investment after deployment, because it shapes the baseline user experience that determines whether the flywheel ever starts spinning.

Connecting Sovereign Pipelines to Commercial Arabic Annotation

Sovereign Arabic LLM programmes represent the large end of the Arabic annotation market, but the pipeline architecture — dialect routing, PDPL compliance, native-speaker SFT and RLHF annotation, cultural safety evaluation — applies at any scale. An enterprise building a Khaleeji Arabic customer service model, a healthtech company building Saudi patient communication AI, or a fintech building Arabic document understanding, requires the same structural approach, scaled to their data volume.

The operational difference between a sovereign programme and a commercial enterprise annotation project is primarily scale, not kind. The dialect routing protocols, annotator qualification frameworks, PDPL documentation requirements, and IAA measurement approaches are the same. What differs is volume (thousands of annotators versus tens), timeline (years versus months), and security clearance requirements.

For Arabic AI teams planning annotation programmes, the sovereign pipeline architecture provides a reference design: if it works at Vision 2030 scale for a national sovereign model, the same structural principles — applied proportionally — will work for your enterprise Arabic LLM project. The key starting point is Arabic NLP annotation infrastructure with genuine dialect routing capability, PDPL compliance, and native-speaker annotator pools qualified by dialect rather than just "Arabic speakers."

Frequently Asked Questions