How MENA foundation-model labs source training data

The five MENA foundation-model programs

Each program differs on dialect coverage, base architecture, training corpus, and intended deployment.

Training data sourcing strategies

Web-scale pretraining corpus

All five labs begin with a web-scale Arabic pretraining corpus. Typical sources:

• Common Crawl — filtered for Arabic language detection + quality
• Arabic Wikipedia — high-quality MSA, ~1.2M articles[^7]
• News + media archives — Al Jazeera, Al Arabiya, Asharq Al-Awsat, others
• Government open data — laws, regulations, official communications
• Books + literature — public domain + licensed corpora
• Religious texts — Quranic + Hadith corpora for cultural grounding
• Scientific + technical — Arabic academic publications, technical translations

Coverage skewed heavily toward MSA. Dialect coverage thin.

Curated SFT (supervised fine-tuning) corpus

After pretraining, models are fine-tuned on instruction-following pairs. Sourcing strategies:

• Translation from English — translate Alpaca-class or ShareGPT-class datasets into Arabic
• Native Arabic instruction generation — write Arabic instructions + responses from scratch
• Hybrid — translate for breadth, native for cultural/dialect fidelity

The translation-only approach produces models that sound like translated English. Native Arabic SFT is the differentiator. This is where curated human annotation matters most.

RLHF preference corpus

For instruction-following + safety, models are aligned via RLHF (Reinforcement Learning from Human Feedback). Sourcing:

• Generate candidate responses
• Human annotators rank candidates
• Train reward model on rankings
• Use reward model to fine-tune base model

Eval set construction

Separately from training, each lab constructs evaluation sets:

• Translated benchmarks (MMLU, HellaSwag, ARC translated to Arabic)
• Native Arabic benchmarks (ArabicMMLU, AraBench, AlGhafa)[^8]
• Dialect-stratified eval sets
• Cultural alignment eval sets (Islamic cultural understanding, gender, politics)

What each program needs

ALLaM (SDAIA)

• Saudi dialect coverage strengthening
• MSA + Saudi government domain
• Saudi cultural alignment eval sets
• PDPL-aware in-Kingdom annotation pipeline
• See ALLaM page

Jais (G42 / Inception)

• UAE Gulf dialect strengthening
• Multilingual Arabic-English code-switching
• Enterprise alignment for UAE corporate use
• See Jais page

Fanar (QCRI)

• Curated quality > quantity thesis
• Islamic cultural domain depth
• Educational + religious alignment
• See Fanar page

Falcon (TII)

• Open-weight model family with Arabic coverage
• Multilingual breadth
• Open-source community engagement
• See Falcon page

Karnak (Egypt AIC)

• Egyptian cultural alignment
• Arabic-Egyptian bilingual code-switching
• See Karnak page

The quality-over-quantity shift

Fanar 2.0 (QCRI, late 2025) was a watershed.[^3] Despite using 8× fewer pre-training tokens than Fanar 1.0, it delivered substantial benchmark improvements (Arabic knowledge, language, dialects, and English).[^6] The signal: curated high-quality Arabic data beats undifferentiated web scrape.

This shift changes the annotation demand profile:

• Pretraining curation matters more — dedup, quality filtering, classifier-based selection
• SFT quality matters more — native Arabic instruction writing by PhD-linguists, not crowdsourced translation
• Dialect stratification matters more — explicit coverage targets per family + sub-family
• Cultural alignment matters more — explicit eval sets, explicit guideline-level cultural sensitivity

This is the work Annota8 was built for.

Where curated workforce changes the model

Three places where curated workforce + PhD-linguist QA materially moves the eval needle:

1. SFT native Arabic instruction writing — translated instructions sound like translations. Natively-written Arabic instructions produce models that sound native. Cairo PhD-linguists with technical domain coverage write 5-10x better instructions than crowdsourced translators.

2. RLHF preference ranking with cultural calibration — preference rankings that don’t account for Arabic cultural context produce misaligned models. Cairo + Riyadh annotators with explicit cultural calibration produce alignment data that holds up.

3. Dialect-stratified eval set construction — eval sets that lump Arabic dialects together hide model weaknesses. Dialect-family-stratified eval sets surface the gaps.

If your foundation-model program is not getting these three from your annotation vendor, you are leaving model quality on the table.

What Annota8 is designed to offer MENA foundation-model labs

Annota8 is being designed for MENA national foundation-model engagements. Capability targets — each scoped per engagement, not pre-certified blanket claims:

• Pretraining-corpus curation (dedup, quality filtering, pretraining-corpus filtering)
• SFT native Arabic instruction writing (Cairo PhD-linguist tier)
• RLHF preference ranking with cultural calibration
• Dialect-stratified eval set construction
• Cultural alignment eval sets
• Sovereign deployment patterns (KSA cloud or on-premise) as design targets
• PDPL-aware design with MENA-resident operator team

Annota8 is in early-stage operations and does not hold formal compliance certifications today. Engagement starts with a controls-mapping conversation with the lab’s compliance team.