PaperlayerView exact revision on arXiv

Source-linked AI summary

Fanar: An Arabic-Centric Multimodal Generative AI Platform

Fanar Team, Ummar Abbas, Mohammad Shahmeer Ahmad, Firoj Alam, Enes Altinisik, Ehsannedin Asgari, Yazan Boshmaf, Sabri Boughorbel, Sanjay Chawla, Shammur Chowdhury, Fahim Dalvi, Kareem Darwish, Nadir Durrani, Mohamed Elfeky, Ahmed Elmagarmid, Mohamed Eltabakh, Masoomali Fatehkia, Anastasios Fragkopoulos, Maram Hasanain, Majd Hawasly, Mus'ab Husaini, Soon-Gyo Jung, Ji Kim Lucas, Walid Magdy, Safa Messaoud, Abubakr Mohamed, Tasnim Mohiuddin, Basel Mousi, Hamdy Mubarak, Ahmad Musleh, Zan Naeem, Mourad Ouzzani, Dorde Popovic, Amin Sadeghi, Husrev Taha Sencar, Mohammed Shinoy, Omar Sinan, Yifan Zhang, Ahmed Ali, Yassine El Kheir, Xiaosong Ma, Chaoyi Ruan

arXiv:2501.13944v1cs.CLcs.AI

TL;DR

Arabic-centric multimodal generative AI remains constrained by limited, culturally aligned Arabic data. Fanar addresses this with concurrently deployed Arabic LLMs and coordinated speech, image, RAG, and attribution services, while its ASR outperforms major speech models on Arabic benchmarks.

  • Problem

    Arabic generative AI is limited by insufficient broad, clean Arabic data and translated content that can miss local cultural nuances.

    Source: §1 · p. 5, §7.1.2 · p. 26

  • Method

    Fanar combines 7B and 9B Arabic LLMs with an orchestrator coordinating multimodal generation, specialized RAG, and attribution services.

    Source: §3 · p. 7, §11 · p. 42

  • Results

    Fanar ASR consistently outperforms Google ASR and Whisper-Largev2 on LAraBench and SADA, while achieving comparable or better results than reported in-domain systems on most datasets.

    Source: §9.1 · p. 38

  • Takeaways & Limitations

    Fanar demonstrates an Arabic-centric platform integrating language, speech, image, retrieval, and factuality capabilities in a single coordinated system.

    Source: §3 · p. 7, §11 · p. 42

    Limited broad, clean Arabic data remains the biggest bottleneck, and human testers reported an approximately 15% dislike rate, mostly due to factuality errors.

    Source: §11 · p. 42

Abstract

from arXiv · show

We present Fanar, a platform for Arabic-centric multimodal generative AI systems, that supports language, speech and image generation tasks. At the heart of Fanar are Fanar Star and Fanar Prime, two highly capable Arabic Large Language Models (LLMs) that are best in the class on well established benchmarks for similar sized models. Fanar Star is a 7B (billion) parameter model that was trained from scratch on nearly 1 trillion clean and deduplicated Arabic, English and Code tokens. Fanar Prime is a 9B parameter model continually trained on the Gemma-2 9B base model on the same 1 trillion token set. Both models are concurrently deployed and designed to address different types of prompts transparently routed through a custom-built orchestrator. The Fanar platform provides many other capabilities including a customized Islamic Retrieval Augmented Generation (RAG) system for handling religious prompts, a Recency RAG for summarizing information about current or recent events that have occurred after the pre-training data cut-off date. The platform provides additional cognitive capabilities including in-house bilingual speech recognition that supports multiple Arabic dialects, voice and image generation that is fine-tuned to better reflect regional characteristics. Finally, Fanar provides an attribution service that can be used to verify the authenticity of fact based generated content. The design, development, and implementation of Fanar was entirely undertaken at Hamad Bin Khalifa University's Qatar Computing Research Institute (QCRI) and was sponsored by Qatar's Ministry of Communications and Information Technology to enable sovereign AI technology development.

1. Introduction

Fanar is introduced as an Arabic-centric multimodal generative AI platform addressing the data, cost, and expertise barriers to high-quality non-English LLMs. Its core comprises two Arabic LLMs trained on nearly 1 trillion tokens, alongside speech, image, specialized RAG, tokenizer, benchmarking, and attribution capabilities.

Motivation

  • Fanar addresses limited Arabic data and the high hardware, monetary, and technical costs of building effective non-English LLMs.Arabic accounts for around 0.5% of web data, while English documents constitute 46% of textual web content.Source: §1 · p. 5

Platform contribution

  • Fanar provides an Arabic-centric generative AI platform spanning text LLMs, speech and image generation, specialized RAG modules, and attribution services.The attribution service is designed to authenticate and correct facts in generated text.Source: §1 · p. 5

Core models

  • Fanar Star and Fanar Prime are 7B and 9B parameter models trained on nearly 1 trillion clean, deduplicated Arabic, English, and Code tokens.Fanar Star was trained from scratch, whereas Fanar Prime was continually pre-trained on Gemma-2 9B; the models work in concert to address Arabic data gaps, especially in technical domains.Source: §1 · p. 5

Arabic adaptations

  • The platform additionally contributes a morphologically aligned Arabic tokenizer and culturally aware benchmarks for evaluating generative AI capabilities.These benchmarks are presented as independently relevant to the generative AI research community.Source: §1 · p. 5

2. Arabic Language

Arabic’s global, cultural, and religious significance coexists with substantial linguistic diversity and structural complexity, motivating language technologies specifically tailored to Arabic. Its diglossia, derivational morphology, script conventions, and limited web representation create distinctive computational challenges.

Scope and significance

  • Arabic is spoken by more than 467 million people across 60 countries, is official or co-official in 25 countries, and is the liturgical language of over 2 billion Muslims.Arabic spans Classical Arabic, Modern Standard Arabic, and diverse regional dialects used across the MENA region and online.Source: §2 · p. 6

Morphology

  • Arabic’s root-and-pattern morphology derives extensive vocabulary from typically three-consonant roots through templates, prefixes, and suffixes that encode nuanced meanings and grammatical features.Inflection marks tense, mood, gender, and number, distinguishing Arabic’s non-linear morphology from predominantly concatenative Indo-European morphology.Source: §2 · p. 6

Diglossia

  • Diglossia complicates Arabic processing because Modern Standard Arabic is used formally but regional dialects are native varieties that differ in phonology, syntax, and lexicon and may be mutually unintelligible.The dialects are also influenced by languages including English, Berber, French, Persian, Turkish, and Aramaic.Source: §2 · p. 6

Script

  • Arabic’s right-to-left script usually omits short vowels, creating challenges for diacritization and semantic disambiguation.The script is derived from the Nabataean alphabet and can represent unrelated languages such as Persian and Urdu.Source: §2 · p. 6

Technology motivation

  • Arabic content represents only 0.5% of online data despite the language’s large speaker population and cultural impact, underscoring the need for culturally and linguistically informed technology development.The passage emphasizes integrating linguistic expertise and anthropological insights into development and evaluation.Source: §2 · p. 7

3. Fanar Platform Services

Fanar is a coordinated suite of Arabic-centric language, speech, image, translation, and retrieval services routed through an orchestrator with safety checks. Its core models and specialized services support culturally aligned, multilingual, domain-controlled generation.

Platform orchestration

  • The orchestrator routes requests from the Chat App or API through safety filtering to appropriate language, speech, translation, image-generation, and other services.Responses are safety-checked for responsible AI guidelines and cultural alignment.Source: §3 · p. 7, Figure 1 · p. 7

Language models

  • Fanar Star is a 7B model trained from scratch on a staged 1-trillion-token Arabic, English, and code corpus, while Fanar Prime is an 8.78B continually pretrained Gemma-2-9B model.A classifier routes non-Islamic, non-STEM prompts to Fanar Star and STEM or reasoning prompts to Fanar Prime; both receive supervised fine-tuning and preference optimization.Source: §3 · pp. 7–8

Speech services

  • Fanar’s Arabic-English speech recognition supports multiple Arabic dialects, non-native accents, and Arabic-English or dialectal code-switching.The platform also provides Arabic and English text-to-speech using a Diffusion Transformer with ConvNeXt V2 for improved text-speech alignment.Source: §3 · p. 8

Image generation

  • The image-generation service fine-tunes Stable Cascade to better reflect Arab and Islamic preferences, including culturally appropriate attire, skin tones, and regional landmarks.Stable Cascade is selected for its smaller latent space and faster fine-tuning and inference.Source: §3 · p. 8

Retrieval Augmented Generation

  • Four RAG systems provide controlled generation through attribution, recency, Islamic, and biography-focused retrieval from external or authoritative sources.They support evidence validation, post-checkpoint information, Islam-related prompts, and accurate biographies of regional and international figures.Source: §3 · p. 8

Translation

  • Fanar provides specialized translation from Modern Standard Arabic or English into Arabic dialects by fine-tuning an existing sequence-to-sequence transformer.The service addresses limited parallel data and builds on QCRI’s 15 years of MSA and dialectal Arabic translation expertise.Source: §3 · p. 8

4. Pre-training Data

Fanar’s pre-training mixture contains approximately 1 trillion cleaned tokens across Arabic, English, and code, with diverse sources, extensive filtering, and deduplication. Machine translation expands Arabic coverage, including MSA and dialectal resources, while specialized evaluation guides system selection.

4.1. Composition of the Fanar Pre-training Data

  • Fanar curates approximately 1 trillion tokens from Arabic, English, and code, drawing on web, scientific, encyclopedic, mathematical, book, news, and programming sources.The final mixture was selected after preprocessing, cleaning, filtering, deduplication, and ablation studies.Source: §4.1 · p. 9, §4 · p. 9

4.1.1. English Data Composition

  • English contributes approximately 513 billion tokens, while code contributes approximately 102 billion tokens from diverse programming and contextual sources.The code subset is about 10% of pre-training data and includes Python, C, C++, Java, Go, JavaScript, Markdown, and GitHub Issues.Source: §4.1.1 · p. 9, §4.1.3 · p. 10

4.1.2. Arabic Data Composition

  • Arabic data comprises 410 billion tokens spanning MSA, Classical Arabic, dialectal content, curated sources, and machine-translated books, STEM papers, and encyclopedic documents.The collection includes web, news, books, Wikipedia, and Arabic poetry to broaden linguistic and cultural coverage.Source: §4.1.2 · p. 10

4.2. Data Curation, Cleaning, and Standardization

  • Arabic data is standardized into JSON records and filtered with Arabic-adapted syntactic, semantic, perplexity, and education-based quality signals before training.The education classifier achieved about 70% validation accuracy, and filtering low scores of 0 and 1 removed approximately 20% of data.Source: §4.2 · pp. 10–11, §4.3 · p. 11, §4.3.1 · p. 11

4.4. Data Deduplication

  • Fanar applies scalable exact-match, approximate-match, URL, inter-dataset, and cross-dataset deduplication to control duplicate content under limited compute and memory.Approximate deduplication uses min-wise LSH with gram size 8, 12 bands, band length 11, and an approximate Jaccard threshold of 0.8.Source: §4.4 · p. 13

4.5. Machine Translation

  • Machine translation expands Arabic training and evaluation resources; English-to-MSA benchmarking selected shaheen, nllb, and S-T5, while dialectal work developed MSA-to-Egyptian and MSA-to-Levantine systems.For dialect translation, NLLB 3.3B surpassed AraT5 and smaller variants after fine-tuning, and human evaluation selected the best system for each dialect despite COMET favoring NLLB for all genres.Source: §4.5 · p. 13, §4.5.1 · p. 14, §4.5.2 · pp. 14–15

5. Tokenization

Fanar introduces MorphBPE, a morphologically aware tokenizer designed to address Arabic’s mismatch with standard BPE while retaining statistical efficiency. Its evaluation shows improved morphological alignment, lower loss, and faster convergence than vanilla BPE.

5.1. Byte-Pair Encoding and its Limitations

  • BPE is widely used for efficient vocabulary management and out-of-vocabulary handling, but its greedy, cross-lingual, and language-unequal character assumptions create limitations.These limitations have motivated variants including BPE dropout, sampling-based BPE, byte-level extensions, and multilingual BPEs.Source: §5.1 · p. 15

5.2. Challenges of Byte-Pair Encoding and Arabic

  • Arabic’s root-and-pattern morphology makes traditional BPE and byte-level approaches poorly suited to capturing meaningful morphological structure.Vanilla BPE can produce semantically unrelated fragments, forcing models to disambiguate components and complicating meaningful embedding learning.Source: §5.2 · p. 16

5.3. Fanar Morphology-based Tokenizer

  • MorphBPE modifies BPE to prevent token merges from crossing morpheme boundaries, preserving Arabic morphology while retaining statistical efficiency.The algorithm initializes character-level vocabulary, applies morphological segmentation, selects frequent byte pairs, and iterates until the target vocabulary size.Source: §5.3 · pp. 16–17

5.4. Tokenization Evaluation

  • The proposed Morphological Alignment Score quantifies how closely tokenizer outputs match morphological segmentations using dynamic-programming pairwise alignment.The score preserves the order of matching tokens and morphemes.Source: §5.4 · p. 17

5.5. Fanar Tokenizer Preprocessing and Training

  • Fanar preprocesses Arabic-script languages by normalizing scripts, removing diacritics from text while retaining them in the vocabulary, and converting data to HuggingFace format.The tokenizer is trained on the complete Arabic development dataset using a vocabulary size aligned with 1024-sized hardware blocks to improve throughput and reduce latency.Source: §5.5 · p. 17
  • The finalized vocabulary combines 32K English tokens with 45K Arabic tokens and totals 75 × 1024 = 76, 800 tokens, reserving capacity for multimodality, diacritics, and special cases.The 45K Arabic vocabulary was selected for optimal morphological alignment, with infrequent merges pruned.Source: §5.5 · p. 17

5.6. Fanar Tokenizer Evaluation Results

  • Morph-BPE demonstrates faster convergence, smaller training loss, and higher morphological alignment than vanilla BPE while maintaining reasonable fertility.Figure 5 compares cross-entropy loss, fertility, and morphological alignment between Morph-BPE and vanilla BPE and other tokenizers.Source: Figure 5 · p. 18, §5.6 · p. 18

6. Modeling and Pre-training

Section 6 defines the architectures, data choices, ablations, and staged pre-training recipes for Fanar Star and Fanar Prime. It reports filtering and mixture trade-offs, curriculum outcomes, optimization settings, and the infrastructure used for training.

6.1. Model Architecture

  • Fanar Star and Fanar Prime are decoder-only Transformers with 7.1B and 8.78B parameters, using distinct from-scratch and continual-training foundations.Star reuses OLMo and Llama architectures, while Prime is built on Gemma-2-9B; their vocabularies contain 76,800 and 128,256 tokens, respectively.Source: §6.1 · p. 18, Table 6 · p. 19

6.2.1. Comparing Data Filtering Strategies

  • About four points on Arabic HellaSwag separate Fanar-filtered from Jais-filtered data, with Fanar filtering sustaining upward performance rather than an earlier plateau.The multi-stage recipe combines perplexity-based filtering and quality classification.Source: Figure 6 · p. 19, §6.2.1 · p. 19

6.2.2. Comparing Data Mixture Composition

  • A 70:20 Arabic-to-English ratio improves Arabic HellaSwag by two points but degrades English performance by six points, motivating dynamic mixing.English tokens remain more prevalent initially, while the Arabic proportion increases toward the end of training.Source: §6.2.2 · pp. 19–20, Figure 7 · p. 20

6.3.1. Training Recipe

  • Fanar Star uses multi-epoch pre-training followed by a high-quality cool-down phase, reaching approximately 3 trillion tokens with steadily decreasing loss and perplexity.Arabic MMLU performance improves progressively, with additional filtering in epoch 3 and a strong downstream boost from cool-down.Source: §6.3.1 · p. 20, Figure 9 · p. 21, Figure 8 · p. 21

6.3.2. Optimization Configuration

  • Fanar Star uses autoregressive training with 4096-token contexts, bfloat16 mixed precision, a 1344-sample global batch, warm-up, cosine annealing, and cool-down to zero learning rate.The optimization configuration uses AdamW and performs critical all-reduce operations in fp32.Source: §6.3.2 · p. 21, Table 7 · p. 21

6.4.1. Training Recipe

  • Fanar Prime continually trains Gemma-2-9B with a pruned 128,256-token vocabulary, balanced 45% Arabic, 45% English, and 10% code data, and a single-epoch-plus-cool-down curriculum.The reported checkpoint is near 600B tokens, while the full 650B-token training run continues.Source: §6.4.1 · p. 22

6.5. Pre-training Infrastructure and Frameworks

  • Training used 168 NVIDIA H100 GPUs across 21 nodes, with OLMo and Dolma supporting Fanar Star and LitGPT enabling Fanar Prime’s continual pre-training.OLMo provided artifact generation, in-loop evaluation, and batch back-tracing for debugging data issues.Source: §6.5 · p. 23

7. Post-Training

Fanar’s post-training combines supervised fine-tuning and preference learning to improve instruction following, Arabic proficiency, harmlessness, and cultural alignment. Its workflow uses filtered bilingual data, staged training, value-aligned synthetic data, preference optimization, annealing, and continuous benchmark and user evaluation.

7. Post-Training

  • Fanar combines supervised fine-tuning and preference learning to produce helpful, harmless, Arabic-proficient, and culturally aligned responses.Training data came from public-source curation and generation by the post-training team using open models and Fanar.Source: §7 · p. 23
  • The post-training strategy introduces sample-level quality validation, multi-stage training, and value-aligned synthetic data generation.These steps target high-quality bilingual data, progressive task reinforcement, and Arabic cultural and religious alignment.Source: §7 · pp. 23–24

7.1.1. Data Curation from Public Sources

  • Approximately 2.5 million bilingual instructions and dialogues across 11 categories were produced through annotated, quality-filtered, translated, and value-relevant public-data curation.The resulting core capability dataset was used during the initial SFT stage.Source: §7.1.1 · pp. 25–26

7.1.2. Synthetic Data Generation

  • Close to a million Arabic and English synthetic samples addressed cultural contextualization and alignment, including 30K responses that mitigated English names and locations in Arabic creative writing.Generation used stronger public models, culturally targeted prompts, rejection sampling, judge scoring, and arena-style pairwise selection.Source: §7.1.2 · pp. 26–27

7.1.3. New Capability Data

  • Targeted datasets expanded coverage of closed-set questions, IslamQ&A, poetry, humor, language tasks, dialectic dialogue, and controversial Islamic topics.Sources included manually expanded prompts, transformed high-quality text, vendor-generated dialectic dialogues, and expert instructional and DPO data.Source: §7.1.3 · pp. 27–28

7.2. Preference Learning

  • Batching preference data improved benchmarks by +1-2%, while balancing Arabic and English samples corrected the model’s tendency to answer in the majority language.DPO used filtered public and on-policy preference data; approximately 250K samples were generated, with around 20% on-policy.Source: §7.2 · p. 28

7.3. Annealing Stage

  • Annealing slightly reduced automated benchmark performance but noticeably reduced user dislike, so it was applied only to Fanar Star while Fanar Prime handled harder queries.The stage uses near-zero learning rates with a curated SFT and DPO subsample to preserve capabilities and rapidly address harmful or culturally misaligned behavior.Source: §7.3 · p. 29

7.5. Collection of User Feedback

  • Continuous benchmarks, standardized user comparisons, and playground feedback guided deployment decisions, with the scratch-trained model reaching an approximately 13% dislike rate.The evaluation suite covered Arabic and English capabilities, while user feedback identified factuality, cultural alignment, grammar, and presentation problems.Source: §7.5 · pp. 29–30, §7.6 · p. 30

8. Evaluation

Fanar Star and Fanar Prime are evaluated against Arabic-aware peers using standard, open-ended, conversational, human, and newly developed culturally and dialectally aware benchmarks. Fanar Prime and Fanar Prime Instruct achieve strong results across Arabic and English evaluations, with particularly strong cultural and dialectal performance.

8.1. Benchmarks

  • The evaluation covers automatic, open-ended and conversational, and human assessments, using comparable Arabic-English baselines across diverse tasks.The framework combines recognized reasoning, reading-comprehension, and question-answering benchmarks with conversational and instruction-following evaluations.Source: §8 · pp. 30–31, §8.1 · p. 31

8.1.1. Automatic evaluation with multi-choice questions

  • Automatic evaluation spans English and Arabic knowledge, reasoning, mathematics, cultural understanding, reading comprehension, and dialectal competence benchmarks.The suite includes MMLU, PIQA, Hellaswag, ARC Challenge, Winogrande, GSM8K, MMMLU, ArabicMMLU, OALL, ACVA, Arab Cultural MCQ, Belebele, Almieyar, and translated dialectal tasks.Source: §8.1.1 · p. 31

8.1.2. Conversation and instruction following evaluations

  • Conversation and instruction-following evaluations cover English MT-Bench, AlpacaEval, and IFEval, plus Arabic MT-Bench, translated AlpacaEval, and 3.5K-chat user-chats, with user-chats judged by GPT-4o.MT-Bench has 80 two-turn questions, AlpacaEval 805 instructions, and IFEval 25 verifiable instructions and 541 prompts. User-chats average 2.7 turns and reach 100 turns.Source: §8.1.2 · p. 32

8.1.3. Human evaluation

  • Human evaluation uses 555 single- and multi-turn dialogues drawn primarily from user messages, with model responses rated by three human evaluators.Evaluators saw only the final prompt and its corresponding response rather than the full dialogue sequence.Source: §8.1.3 · p. 32

8.2. Baselines

  • Baselines are capable Arabic-English or multilingual models in Fanar’s size range, including Jais, OLMo, Allam, Qwen 2.5, and Gemma 2.Source: §8.2 · pp. 32–33

8.3.1. Base models

  • Results report accuracy or normalized accuracy for multi-choice tasks, exact match with flexible extraction for GSM8K, and scores out of 10 for LLM-as-a-judge benchmarks.Tables 10–13 separate Arabic and English results for base and instruction-tuned models and mark the best and second-best scores.Source: §8.3.1 · p. 33, §8.3.2 · p. 33
  • Fanar Prime achieves the best result on most Arabic benchmarks, including cultural benchmarks, while outperforming Gemma-2-9b-it on most English benchmarks.Its only stated Arabic exception is dialectal PIQA; in English, it reaches the best or second-best score overall.Source: §8.3.1 · p. 33, Table 10 · p. 34, Table 11 · p. 34

8.3.2. Instruction-tuned models

  • Fanar Prime Instruct ranks among the top two in nearly all Arabic evaluations and provides competitive English performance, with especially strong results on cultural and dialectal tasks.It shows superior performance on most standard Arabic benchmarks and competitive open-ended results judged by LLMs and humans.Source: §8.3.2 · pp. 33–34, Table 12 · p. 35, Table 13 · p. 36

9. Integrating Multimodal Support

Fanar integrates speech input/output and culturally aligned image generation to support Arabic-centric multimodal interaction. Its speech system emphasizes inclusive multilingual and dialectal coverage, while its image system addresses cultural knowledge and preference biases through targeted fine-tuning.

9. Integrating Multimodal Support

  • Fanar supports speech interaction in English and Arabic, including multiple Arabic dialects, and generates culturally aligned imagery appropriate for Arabic-speaking regions.Users can provide speech input and receive model-generated responses in audio format, while image generation targets regional cultural alignment.Source: §9 · p. 37

9.1. Speech Modality

  • 7.5% WER on Libri-Clean and 12% on Libri-Other, while Fanar ASR consistently outperforms Google ASR and Whisper-Largev2 on LAraBench and SADA datasets.Fanar ASR also achieves significantly better or comparable results to reported in-domain systems on most evaluated datasets, despite allocating only 25% of vocabulary tokens to English.Source: Figure 15 · p. 38, §9.1 · p. 38

9.2. Image Modality

  • Fanar’s image system improves cultural nuance by fine-tuning Stable Cascade with Arab-world visual concepts and preference optimization; its averaged model performs better than each individual model.The taxonomy contains more than 5000 concepts and yields about 100,000 high-quality images, while direct comparisons use identical initializations.Source: §9.2 · pp. 40–41

10. Retrieval Augmented Generation

Fanar uses multiple Retrieval Augmented Generation systems to improve domain-specific accuracy across Islamic, biographical, recent-information, and fact-attribution prompts. These systems combine retrieval, reranking or search, response generation, and— for attribution—post-generation fact-checking and revision.

10. Retrieval Augmented Generation

  • Fanar applies RAG to Islamic content, biographical information, recent information, and attribution of fact-related prompts.These systems target specific domains to improve the accuracy of model-generated content.Source: §10 · p. 41

10.1. Islamic RAG

  • Islamic RAG stores curated literature in Milvus, retrieves twenty threshold-matched documents, reranks four, and constructs a context-optimized prompt for Fanar.Documents are chunked into up to 2048-token segments with 50-token overlap at semantic boundaries, while system hyperparameters and thresholds are tuned on domain-specific benchmarks.Source: §10.1 · p. 41

10.2. Recency and Biography RAG

  • An orchestrator classifies recent-information and biographical prompts, routes them to specialized RAG systems, and has Fanar Prime search, summarize, and cite the results.Fanar Prime rewrites each routed prompt as a search query, accesses relevant information through a search engine, and generates the response with the included source.Source: §10.2 · p. 42

10.3. Attribution RAG

  • Fanar’s optional attribution service uses post-generation RAG to retrieve evidence, fact-check and revise responses, and add references without modifying the underlying models.The service uses an LLM to revise responses from top-k relevant fact documents, preserving properties such as style and structure; this implements a research-then-revise methodology (Gao et al., 2022).Source: §10.3 · p. 42

11. Discussion and Future Plans

Fanar’s first release combines two concurrently operated Arabic-centric LLMs with multimodal services, RAG systems, attribution, orchestration, and safety filtering. The discussion identifies data scarcity, factuality, non-agentic orchestration, non-unified multimodality, enterprise adoption, and sustainable data access as priorities for future development.

First release

  • Fanar’s first release combines Fanar Star and Fanar Prime with speech and image generation, four RAG systems, attribution, orchestration, and prompt- and response-level safety filters.Fanar Star is a 7B model trained from scratch on nearly 1 trillion tokens, while Fanar Prime continually trains Gemma-2 9B on the same token set.Source: §11 · p. 42

1. Data and factuality

  • Arabic data scarcity is the biggest bottleneck, while coupling Fanar Star with Fanar Prime provides a pragmatic balance between English-data coverage and Arabic characteristics.Arabic-centric benchmark results are described as robust, but human testers’ dislike rate is around 15%, mostly because of factuality errors, motivating larger models, extensive RAG, and attribution.Source: §11 · pp. 42–43

2. Agency and reasoning

  • Fanar’s current orchestrator is reactive rather than agentic, so the next generation will add reasoning, verification, tool calling, and enhanced test-time computation for difficult tasks.The proposed test-time approach generates multiple responses and uses an external reward model to select the best answer (Snell et al., 2024).Source: §11 · p. 43

3. Unified multimodality

  • Future Fanar versions will unify speech, image, and text generation within one generative model and add culturally appropriate, efficient video generation.The current multimodal capabilities are not intrinsic to an autoregressive model with combined image and speech tokens.Source: §11 · p. 43

4. Applications and adoption

  • Fanar’s initial applications span education, news summarization, and government-information chatbots, while future evolution targets enterprise integration driven by real productivity benefits.The stated target includes both government and private-sector use.Source: §11 · p. 43

5. Data sustainability

  • TokenX uses blockchain and LLM attribution to pay publishers when subscribed providers use their content, aiming to preserve open access and expand Arabic content creation.The motivation includes declining publisher traffic and revenue, emerging copyright-payment questions, and restrictions affecting over 5% of Web data and 25% of high-quality Web sources (Longpre et al., 2024).Source: §11 · p. 43

A. Contributions · B. Arabic-Speaking Map and Language Statistics

The paper organizes Fanar’s development into multiple work packages spanning data, model training, evaluation, applications, deployment, and leadership. It also presents the Arabic-speaking regions and contrasts population with web-content availability across major languages.

A. Contributions

  • Fanar’s authors are assigned across distinct work packages and sub-projects, with names within each subgroup ordered alphabetically.The listed packages cover data collection, pre-training, post-training, benchmarking, applications, engineering, infrastructure, and scientific leadership.Source: A. Contributions · p. 44
  • The contribution structure includes separate leads for Fanar Star and Fanar Prime pre-training, post-training, benchmarking, and the morphology-based tokenizer.Additional listed responsibilities include machine translation, image generation, speech, Islamic RAG, attribution, and external user study and system testing.Source: A. Contributions · p. 44
  • The project assigns dedicated roles for applications, engineering and deployment, infrastructure, and scientific and management leadership.These assignments are listed alongside specialized technical and evaluation work packages.Source: A. Contributions · p. 44

B. Arabic-Speaking Map and Language Statistics

C. In-house Benchmarks · C.1. Almieyar: capability-based benchmarking

Almieyar introduces capability-focused benchmarks centered on Arabic language and culture, addressing limitations of entangled evaluations through structured linguistic assessment and an extensible dataset-generation recipe.

C.1. Almieyar: capability-based benchmarking

  • Almieyar provides capability-focused Arabic benchmarks designed to reveal specific LLM strengths and weaknesses beyond entangled multi-skill evaluations.The broader family targets multiple LLM-related capabilities, while this section focuses on Almieyar-Language.Source: §C.1 · p. 46
  • The Almieyar dataset is created through LLM template-based generation, native-speaker review for linguistic and cultural accuracy, and professional linguistic validation.These three steps form the Almieyar-Recipe for dataset construction.Source: §C.1 · p. 46
  • Almieyar-Language evaluates Arabic linguistic understanding through a structured framework that can extend to other languages and addresses gaps in English-only benchmarks such as Holmes (Waldis et al., 2024).It includes linguistic categories omitted by Holmes, including phonology, and is grounded in linguistic capabilities.Source: §C.1 · p. 46
  • The taxonomy covers five linguistic layers—phonology, morphology, syntax, semantics, and pragmatics—with special attention to Arabic characteristics described in.The taxonomy is summarized in Figure 18.Source: §C.1 · p. 46
  • Almieyar-Language incorporates contributions from speakers of 16 Arabic dialects and emphasizes dialect-specific lexical semantics and pragmatics.This design addresses Arabic’s dialectal diversity.Source: §C.1 · p. 46
  • Fanar is evaluated against prominent models across the five linguistic layers using approximately one thousand high-quality questions drawn from diverse linguistic categories.The evaluation uses the Almieyar-Language dataset to provide detailed performance insights.Source: §C.1 · p. 46
  • Almieyar-Overall adds manually reviewed Arabic questions covering knowledge, reasoning, alignment, and robustness for rapid feedback after SFT training.Its hierarchical structure is shown in Figure 19; Culture and Multimodality segments are reserved for separate reporting.Source: §C.1 · p. 46

C.2. Arab Cultural MCQ · D. LLM Security and Safety

The paper introduces a validated 1K-question Arab cultural multiple-choice dataset and describes aiXamine, a security and safety evaluation suite applied to Fanar Star and other open-source models. Together, these contributions address culturally relevant Arabic benchmarking and systematic assessment of LLM vulnerabilities.

C.2. Arab Cultural MCQ

  • The Arab Cultural MCQ dataset contains 1K questions curated from 7.5K GPT-4o-generated questions grounded in high-quality Arabic-culture web content. [17]The source pages covered Arabic cultural issues, values, and nuances.Source: §C.2 · p. 46
  • Three-stage quality control used expert factuality and language review, relevance and difficulty filtering, and manual removal of semantically redundant questions.The final deduplication stage used bge-multilingual-gemma2 embeddings, which were reported to perform well in Arabic (Chen et al., 2024a).Source: §C.2 · pp. 46–47
  • The dataset targets culturally meaningful questions with non-obvious answers and meaningful choices rather than merely testing easy factual recall.Reviewers selected questions relevant to Arabic culture and removed questions that were easy, obvious, or poorly constructed.Source: §C.2 · p. 47

D. LLM Security and Safety

  • aiXamine evaluates LLM security and safety through more than 30 tests spanning data vulnerabilities, model robustness, privacy, bias, hallucination, refusal behavior, and content safety.Its data tests include poisoned samples, backdoors, noise, bias, and potential copyright violations; model tests also cover out-of-distribution behavior and watermark verification.Source: D. LLM Security and Safety · p. 50
  • The evaluation suite assesses harmful or inappropriate text and code generation alongside adversarial attacks, unfair behavior, hallucinations, and inappropriate refusal of innocuous requests.These tests are designed to evaluate both data safety and model safety.Source: D. LLM Security and Safety · p. 50
  • Table 14 reports aiXamine scores for selected open-source models and compares them with Fanar Star.The supplied passage identifies the comparison scope but does not provide the table’s individual scores.Source: D. LLM Security and Safety · p. 50, Table 14 · p. 50

Related papers

Loading 2501.13944v1…