PaperlayerView exact revision on arXiv

Source-linked AI summary

Jais 2: A Family of Arabic-Centric Open Large Language Models

Mohamed Anwar, Abed Alhakim Freihat, George Ibrahim, Mostafa Awad, Abdelrahman Sadallah, Gurpreet Gosal, Gokulakrishnan Ramakrishnan, Sarath Chandran, Biswajit Mishra, Rituraj Joshi, Ahmed Frikha, Etienne Goffinet, Abhishek Maiti, Ali El Filali, Sarah AlBarri, Samujjwal Ghosh, Rahul Pal, Parvez Mullah, Awantika Shukla, Sajid siddiki, Samta Kamboj, Onkar Pandit, Sunil Kumar Sahu, AbdelRahman Elbadawy, Amr Mohamed, Ahmad Chamma, Evan Dufraisse, Abdelaziz Bounhar, Dani Bouch, Hadi Abdine, Guokan Shang, Fajri Koto, Yuxia Wang, Zhuohan Xie, Ali Mekky, Rania Elbadry, Sarfraz Ahmad, Momina Ahsan, Omar El Herraoui, Daniil Orel, Hasan Iqbal, Kareem Elzeky, Mervat Abassy, Kareem Elozeiri, Saadeldine Eletter, Farah Atif, Nurdaulet Mukhituly, Haonan Li, Xudong Han, Aaryamonvikram Singh, Zainul Abedien Ahmed Quraishi, Neha Sengupta, Larry Murray, Avraham Sheinin, Joel Hestness, Natalia Vassilieva, Hector Xuguang Ren, Zhengzhong Liu, Michalis Vazirgiannis, Preslav Nakov

arXiv:2608.13580v1cs.CLcs.AI

TL;DR

Arabic remains underrepresented in global training corpora, and general-purpose models show partial proficiency in dialectal and culturally nuanced contexts. Jais 2 presents an Arabic-centric model family and achieves leading results among evaluated open models on OALL2 and AraGen, as well as in translation and summarization.

  • Problem

    Arabic remains underrepresented in global training corpora, while general-purpose models show partial proficiency, especially in dialectal and culturally nuanced contexts.

    Source: §1.1 · p. 4

  • Method

    The paper presents an Arabic-centric LLM family comprising a 70B model trained entirely from scratch and a competitive 8B variant.

    Source: §11 · p. 62

  • Results

    Jais 2 achieves leading results among evaluated open models on OALL2 and AraGen, and leads in translation and summarization.

    Source: §1.2 · p. 5

  • Takeaways & Limitations

    The models’ open release under a commercially permissive license is intended to foster collaboration, reproducibility, and continued progress in Arabic AI research.

    Source: §11 · p. 62

    Despite efforts to reduce dataset bias, Jais 2 may still exhibit biases, so users should assess them for their use cases before deployment.

    Source: §12.3 · p. 63

Abstract

from arXiv · show

Jais 2 is a family of Arabic-centric large language models developed jointly by MBZUAI, Cerebras, and Inception, designed to advance Arabic-centric language modeling, with strong performance across the Arabic and culturally grounded benchmarks evaluated in this report. The family includes, to our knowledge, the largest open Arabic-centric LLM trained from scratch at 70B parameters, and a competitive 8B-parameter variant among the evaluated open models. A custom Arabic-centric vocabulary enables efficient training and inference. In addition, an optimized architecture and training recipe yield highly compute-efficient training. With a substantially smaller token budget than comparable models, Jais 2 achieves strong Arabic performance on the benchmarks considered in this report and competitive English results. The models obtain leading results among the evaluated open models on OALL2 and AraGen. They also perform strongly on several culturally grounded Arabic benchmarks, including poetry, religion, cuisine, and dream interpretation, as well as in general tasks such as translation and summarization. We release the models in HuggingFace under a commercially permissive license. Jais 2 70B is also released as a chat app on the Web, iOS, and Android; it runs on Cerebras hardware, delivering up to 2,000 tokens per second, and enabling high-throughput Arabic-centric chat serving in our deployment setting. By uniting scale, linguistic diversity, cultural fidelity, openness, and speed, Jais 2 provides an open-weight foundation intended to support further research and development in Arabic-centric LLMs.

1 Introduction

Jais 2 is an open Arabic-centric LLM family trained from scratch in 8B and 70B variants, designed for broad Arabic coverage, cultural grounding, and competitive English performance. It reports leading results among evaluated open models on OALL2 and AraGen, and combines efficient training with open release and chat deployment.

1 Introduction

  • Jais 2 is trained from scratch in 8B and 70B variants using domain-diverse Arabic data, dialect and script coverage including Arabizi, and bilingual training for competitive English performance.The report describes the 70B model as among the largest open Arabic-centric models trained entirely from scratch and the 8B model as competitive among similarly sized open models.Source: §1 · p. 4, §1.2 · p. 5
  • A custom 150K-token Arabic-optimized vocabulary, an expanded 8× feedforward filter ratio, ReLU2, and µP support efficient training, while multi-stage preference alignment targets instruction-following and safety.The training curriculum includes continual pretraining, supervised fine-tuning, DPO, and GRPO.Source: §1 · p. 4, §1.2 · p. 5

1.1 Motivation and Context

  • 600B+ curated Arabic tokens and 1.6T tokens of web, math, and code data support native coverage of Modern Standard Arabic and regional dialects alongside culturally grounded post-training.The post-training covers poetry, religion, cuisine, dream interpretation, translation, and summarization.Source: §1.1 · pp. 4–5

1.2 Summary of Contributions

  • Jais 2 leads among evaluated open models on OALL2 and AraGen, and on translation and summarization, while performing strongly across Arabic dialects, cultural domains, and English tasks.Its culturally grounded domains include poetry, cuisine, and dream interpretation.Source: §1 · p. 4, §1.2 · p. 5
  • Both model variants are released as open weights under a commercially permissive license, and Jais 2 70B is available through web and mobile chat apps.In the described deployment setting, the Cerebras-powered chat app delivers up to 2,000 tokens per second.Source: §1 · p. 4, §1.2 · p. 5

2 Related Work

Prior Arabic-language models and resources have advanced bilingual modeling and cultural alignment, but dialectal diversity and culturally grounded evaluation remain gaps; Jais-2 addresses them through broad Arabic instruction data and Arabic-native evaluation.

Arabic Language Models

  • Arabic models span bilingual Jais and Jais-adapted systems and culturally aligned AceGPT, ALLaM, and Fanar, yet dialect models such as Nile-Chat and Atlas-Chat remain separate (Sengupta et al., 2023; Gosal et al., 2024; Huang et al., 2024; Bari et al., 2025; Team et al., 2025a; Shang et al., 2025a,b).The passage notes that general-purpose models have largely overlooked Arabic dialects.Source: §2 · p. 8

Arabic Training Data

  • Instruction tuning draws on multilingual mixtures, including Arabic-extended Aya and xP3, while Arabic pipelines combine native supervision with translated English instructions (Sengupta et al., 2023; Gosal et al., 2024; Bari et al., 2025; Huang et al., 2024; Team et al., 2025a).Instruction tuning uses prompt–response pairs and preference signals such as DPO.Source: §2 · p. 8
  • Jais-2’s bilingual instruction tuning uses re-processed Arabic spanning MSA, 17 dialects, Arabizi, poetry, religion, and science, plus culturally grounded instruction sets for domains including dream interpretation, Islamic QA, and poetry.The strategy explicitly targets dialect diversity and script variation.Source: §2 · p. 8

Arabic Evaluation Benchmarks

  • Existing evaluations cover knowledge and reasoning, instruction following and generation, and safety and bias, but offer limited dialect, script, and cultural coverage; Jais-2 adds Arabic-native domain tasks across linguistic contexts.The framework complements standard benchmarks to assess cultural grounding, long-form reasoning, and safety.Source: §2 · p. 8

3 Model

Jais 2 combines a decoder-only Transformer with empirically selected architectural and training choices, specialized training infrastructure, and a custom multilingual BPE tokenizer.

3.1 Model Architecture

  • Jais 2 adopts a decoder-only Transformer, using ablation-guided choices including decoupled embeddings, wide FFNs, ReLU2, and RoPE.The width-to-depth ratio is kept near 100, and scaling experiments use 31 tokens per parameter with an English:Arabic:Code mixture of 2:1:0.4. The FFN filter-to-hidden-size ratio is 8; µP supports wider intermediate sizes, while ReLU2 increases activation sparsity. RoPE outperformed ALiBi at training and longer zero-shot contexts in controlled ablations. Training uses longer context than Jais 1’s 2,048 tokens, with scaling-law experiments comparing 2,048 and 8,192; attention masking across packed-document boundaries yielded no pretraining compute-efficiency gains.Source: §3.1 · pp. 9–10

3.2 Training Hyper-parameter Values

  • µP enabled zero-shot transfer of tuned hyperparameters from a 100-million-parameter proxy model to larger Jais 2 models.The proxy had hidden size 256, depth 68, and was trained for 20 tokens per parameter; the search covered learning rate, initialization, embedding and unembedding scales, and layer-type-specific scales. Base initialization standard deviation was 0.035 and base learning rate 0.0248; output-layer scales were adjusted by depth^0.5. Batch sizes were 408 for 8B and 960 for 70B initially, increasing to 1,920 for 70B after gradient-noise-scale analysis.Source: §3.2 · p. 10
  • Jais 2 uses AdamW with β1 = 0.9, β2 = 0.95, and ϵ = 10^-15 to preserve effective layer-wise adaptive learning rates.The smaller ϵ addresses layers where the squared-gradient EMA scale is comparable to ϵ, allowing the denominator to remain governed by the gradient estimate rather than the constant. Training uses weight decay 0.1 and a two-phase schedule that warms up linearly to 0.0248, then decays linearly to zero.Source: §3.2 · p. 10

3.3 Training Infrastructure

  • Final Jais 2 training and fine-tuning ran on up to 64 CS-2 systems across Cerebras CG-1 and CG-2 clusters.The clusters use weight streaming and data parallelism; scaling to 64 systems delivered close to 64× speedup with minimal overhead. The last GRPO stage instead ran on A100- and H100-based GPU nodes.Source: §3.3 · p. 11

3.4 Tokenizer

  • Jais 2 uses a custom BPE tokenizer with a 150,272-token vocabulary, trained on multilingual and programming-language text.Weighted sampling emphasized English and Arabic, with smaller proportions of languages such as French and Hindi to support broader coverage and potential future adaptation without retraining the tokenizer.Source: §3.4 · p. 11

4 Pretraining

Jais 2 pretraining combines a two-stage curriculum with curated Arabic data, selected English web data, and targeted data mixing and upsampling. The approach uses broad web data first, then focused high-quality data to address remaining capability gaps.

4 Pretraining

  • The two-stage curriculum first builds general-purpose linguistic and world knowledge from diverse web data, then uses high-quality domain-specific data to target weaknesses.Stage 1 often consumes over 90% of total FLOPs; Stage 2 uses reduced learning rates and around 5–10% of total FLOPs to refine capabilities such as reasoning, code generation, or multilingual understanding.Source: §4 · pp. 11–12

4.1 Arabic Pretraining Data

  • Jais 2’s Arabic preprocessing normalized text, retained documents unless noisy content was the majority, cleaned acceptable noise, and removed JavaScript fragments, long URLs, and documents left too short or empty.The updated pipeline relaxed filtering to retain more clean Arabic text and added normalization, including converting encoded religious-expression symbols to explicit text forms.Source: §4.1 · pp. 12–13
  • The Arabic corpus included 17 dialect variants totaling approximately 600 million tokens, alongside poetry, literature, religious, and scientific material.Religious data underwent expert manual review and refinement and totaled 26 B tokens; the combined reprocessed and specialized Arabic pretraining corpus contained 624B tokens.Source: §4.1 · p. 13

4.2 English Pretraining Data

  • Corpus 2 was best for English, while Corpus 3 was best for Arabic; Corpus 3 was selected for strong larger-scale performance and better cross-lingual transfer to Arabic.Comparisons used 12 runs across three corpora and four model sizes, trained with 20 TPP and identical hyperparameters; accuracy averaged 0-shot evaluations across English and Arabic tasks.Source: §4.2 · pp. 13–14, Table 3 · p. 14

4.3 Pretraining Data Mixing

  • Stage 1 combined general English web data with high-quality Arabic cultural and regional topics, mathematics, and coding; Stage 2 upsampled focused web, curated domain, math, and reasoning data.The second annealing stage used 10% of training FLOPs to address remaining deficiencies.Source: §4.3 · p. 14

4.3.1 Stage 1: Pretraining

  • The final Stage 1 mix was 28% Arabic, 40% Web, 20% Math, 10% Code, and 2% Synthetic data.At 6.7B scale, an Arabic:English ratio of 1:1.5 gave better Arabic accuracy while using all Arabic tokens to support cross-lingual transfer, motivating the increased Arabic fraction.Source: §4.3.1 · p. 15

4.3.2 Stage 2: Annealing

  • Domain upsampling increases the share of domain-specific datasets toward the last training phase, after enough FLOPs have been run to measure signal on difficult benchmarks.Jais 2’s approach differs from the original by not upsampling at the cost of the original datasets.Source: §4.3.2 · p. 15

4.3.3 Upsampling Data Mix Strategy

  • During the annealing phase, targeted-domain upsampling improved performance, with 5–9% average-score gains on Arabic, English, and math tasks; US Mix 4 improved English and Arabic over baseline without significant Math regression.Gains were strongest in specifically upsampled domains and robust across model sizes.Source: §4.3.3 · p. 15

5 Post-Training

Jais 2 post-training refines pretrained capabilities and alignment through continual pretraining, instruction fine-tuning, and preference alignment, targeting linguistic grounding, instruction following, and alignment with human values and cultural context.

Continual Pretraining (CPT)

  • CPT trains the base model for two epochs on curated Arabic data mixed with replay from its original pretraining corpus.It targets key domains and weaker areas identified after initial IFT.Source: §5 · p. 16

Instruction Fine-Tuning (IFT)

  • IFT uses over 20 million instruction–completion pairs across three epochs to align model behavior with human intent.The data includes rewritten Jais 1 examples, open Arabic resources, Arabic and English synthetic data, culturally rich tasks, and translated English datasets where Arabic data was scarce.Source: §5 · p. 16

Preference Alignment

  • Preference alignment uses DPO to steer outputs toward human preferences and ethical principles and teach safe, helpful assistant behavior.Source: §5 · p. 16

Post-Training

  • Post-training combines three stages to strengthen Jais 2’s linguistic grounding, instruction following, and alignment with human values and cultural context.The stages are continual pretraining, instruction fine-tuning, and preference alignment.Source: §5 · p. 16

6 Continual Pretraining

Jais 2’s base model is continually pretrained on a newly constructed corpus to strengthen specialized-domain capabilities and address weaknesses observed after initial IFT experiments.

6 Continual Pretraining

  • Continual pretraining uses two epochs on a newly constructed corpus, with 50% replay from the original pretraining dataset.The stage aims to enhance knowledge and capabilities in intended specialization domains and strengthen areas found suboptimal after initial IFT experiments.Source: §6 · p. 16
  • Specialized curated Arabic datasets support specialization, while synthetic English and Arabic data target weaker areas through textbook-style content, explanation-enriched MCQs, and math datasets.The textbook-style content follows an approach inspired by Neema et al. (2025).Source: §6 · p. 16

7 Instruction Fine-Tuning

This section describes Jais 2’s instruction-fine-tuning data and task designs, combining broad Arabic and English resources with dialect-specific and culturally focused datasets across translation, dialect identification, dream interpretation, and summarization.

7 Instruction Fine-Tuning

  • IFT addresses the mismatch between next-token pretraining and instruction following by combining broad resources with culturally focused Arabic poetry and dream-interpretation data.The authors also experimented with GRPO preference alignment, but did not use it in the released models.Source: §7 · p. 16

7.1 General IFT: Standard Arabic and English

  • The general IFT corpus contains over 20M diverse data points, including enhanced Jais 1 SFT data, public and synthetic datasets, and Arabic culture-centric collections.Synthetic data spans Arabic and English and targets instruction following, dialogue, safety, reasoning, STEM, translation, grammar, and sentiment tasks.Source: §7.1 · p. 17
  • Synthetic training examples were decontaminated by checking each against benchmark test examples, and a few thousand system-prompt examples taught adherence to system instructions.The decontamination process removed examples judged contaminated by the LLM-decontaminator.Source: §7.1 · p. 17
  • The authors found that scaling IFT alone led instruction-following performance to plateau, indicating it was insufficient for robust complex instruction following.Source: §7.1 · p. 17

7.2 General IFT: Dialectal Arabic

  • Because Arabic speakers commonly prefer local dialects for everyday interaction, Jais 2’s general dialect IFT datasets target Darija and Egyptian Arabic.The datasets were systematically collected, annotated, and validated to represent each dialect’s lexical, morphological, and syntactic features.Source: §7.2 · pp. 17–18

7.2.1 Darija-SFT-Mixture

  • The Darija-SFT-Mixture contains 458K instructions covering translation, transliteration, summarization, and synthetic social-media tasks such as replies, explanations, and safe responses.It also adapts high-quality instruction-tuning data from TULU-v2 through machine translation.Source: §7.2.1 · p. 18

7.2.2 Egyptian-SFT-Mixture

  • The Egyptian-SFT-Mixture contains 1.85M instructions, with data for translation between Egyptian Arabic, MSA, and English, transliteration, and multi-turn tasks.The curation draws on Aya Collection resources, Egyptian Wikipedia, and translated TULU v2 and v3 data.Source: §7.2.2 · p. 18

7.3 Task: Dialectal Arabic Translation

  • Dialectal translation training uses a cleaned corpus of 612,916 translation pairs and 15,731,037 tokens, with balanced MSA–dialect, dialect–MSA, and Arabic–English directions.Template injection turns parallel pairs into varied Arabic and English instruction–response examples, with dialect metadata mapping each pair to an appropriate prompt.Source: §7.3 · pp. 19, 22
  • Over 20 parallel sources were unified into an instruction-following dataset for zero-shot dialectal translation, with source–target pairs used in both directions between MSA and English.The directionality balancing process generated separate instructions for X → Y and Y → X.Source: §7.3 · p. 22

7.4 Task: Arabic Dialect Identification

  • The dialect-identification IFT training data contains 624K examples covering 15 dialects and languages; sentence-label samples were reformatted as instruction–response pairs, with at least ten distinct prompt templates created for each dialect.Source: §7.4 · p. 23

7.5 Task: Arabic Cuisine

  • The Arabic cuisine work introduces an IFT dataset and a recipe-grounded benchmark of 1,597 human-validated five-way questions with plausible hard distractors.Benchmark dimensions cover culinary knowledge and execution as well as cultural alignment; recipe-level decontamination was used to reduce train–test leakage.Source: §7.5 · p. 25

7.6 Task: Arabic Poetry

  • The poetry IFT data supports both analysis and generation, training capabilities in metadata reasoning, coherent poem generation, and understanding Arabic verse’s stylistic and linguistic nuances.Poems were enriched with keyword and keyphrase metadata, deduplicated, and removed from training if they appeared in the FannOrFlop test benchmark.Source: §7.6 · p. 26

7.7 Task: Islamic Question-Answering

  • The Islamic QA IFT dataset contains 150,890 instruction–response examples, each incorporating an ethical disclaimer to communicate the purpose and limitations of AI-generated answers.The disclaimer addresses the sensitive nature of religious questions.Source: §7.7 · p. 27

7.8 Task: Dream Interpretation

  • To address the limited study of culturally grounded dream interpretation, the authors construct a bilingual multiple-choice benchmark from Arabic and Western cultural sources.It evaluates symbolic meaning, culturally appropriate interpretations, and discrimination among plausible alternatives in Arabic dream contexts.Source: §7.8 · p. 28

7.9 Task: Summarization

  • The summarization corpus combines human-written and synthetic sources across Arabic genres, dialects, and monolingual and cross-lingual tasks, with 540K training examples and 25K evaluation examples.It contains 178M training tokens and 9.5M evaluation tokens; prompt templates cover Arabic, English-to-Arabic, and Arabic-to-English summarization.Source: §7.9 · pp. 28–29

8 Preference Alignment

Jais 2 uses DPO trained on curated preference pairs to align model behavior, improving English and Arabic instruction-following and Vicuna evaluation win rate. Its data pipeline combines regenerated instructions, model-generated responses, and frontier-LLM judging.

8 Preference Alignment

  • DPO directly optimizes model parameters from preference data without an external reward model.It increases the preferred response’s relative likelihood while the reference-policy log-ratio implicitly regularizes the optimized policy.Source: §8 · p. 30
  • DPO compares preferred and less-preferred responses using the Bradley–Terry model and their log-likelihood ratios under the current and reference policies.The reference policy is typically the SFT model, and β controls the strength of the update.Source: §8 · p. 30
  • The team curated over 200k chosen–rejected preference pairs across general conversation, Arabic, math, and instruction following using self-play and frontier-LLM judging.Seed prompts from public preference collections were regenerated as high-quality instructions; the model generated responses, which a frontier LLM judged and replaced with preferred responses.Source: §8 · p. 30, Figure 7 · p. 30
  • DPO improved Jais 2’s English and Arabic instruction-following performance and its win rate in Vicuna evaluations.The selected learning rate, batch size, and β were 4.0e-5, 160, and 0.1, respectively.Source: §8 · p. 30

9 Evaluation

The evaluation assesses Jais 2 across Arabic multiple-choice and generative benchmarks, translation, dialect identification, culturally grounded domains, instruction-following, summarization, and English capabilities. It reports performance at both model sizes, notes benchmark-specific gaps, and includes Arabic-specific instruction-following tests with strict and loose scoring criteria.

9.1 MCQ Evaluation: Open Arabic Leaderboard 2 (OALL2)

  • 72.40% on OALL2 is Jais 2 8B’s highest macro-average among models with at most 13B parameters; Jais 2 70B leads models above 13B with 79.36%.The 8B model also leads its size range on ArabicMMLU, while the 70B model has the highest accuracy on AlGhafa, ArabicMMLU, AraTrust, and MadinahQA.Source: Table 10 · p. 31, §9.1 · p. 32

9.2 Generative Evaluation: AraGen

  • Jais 2 achieves the highest overall AraGen scores among the compared open multilingual and Arabic-centric models at both ≤13B and >13B scales.AraGen uses LLM-as-a-judge and the 3C3H measure, assessing Correctness, Completeness, Conciseness, Helpfulness, Harmlessness, and Honesty; its results suggest practical utility for Arabic interactions.Source: §9.2 · pp. 32–33, Table 11 · p. 32

9.3 Arabic Translation

  • On general Arabic–English translation, Jais 2 70B achieves the highest overall performance, while Jais 2 8B performs strongly among mid-sized Arabic-centric and multilingual models.The six datasets span classical, formal, colloquial, and cross-domain translation; across translation benchmarks, the 8B model is among the strongest in its class and the 70B model leads evaluated open models.Source: §9.3 · p. 33, Summary · p. 33, Table 12 · p. 34
  • Across dialect translation directions, Jais 2 70B achieves the highest or near-highest BLEU scores, while Jais 2 8B ranks among the strongest models in its size range.The 70B model is strongest among open systems for translation into dialects and from dialects to MSA or English; gains are especially pronounced for dialect-to-MSA and dialect-to-English translation.Source: §9.3 · p. 33, Table 13 · p. 35, Table 14 · p. 36

9.4 Arabic Dialect Identification

  • Jais 2 70B ranks first across nearly all evaluated dialect categories and has the highest average accuracy on both MADAR and QADI.Jais 2 8B averages 46.03% accuracy on MADAR and 44.24% on QADI, placing it among the top mid-sized Arabic-centric models; QADI uses noisy, code-switched social-media text.Source: §9.4 · p. 39, Table 16 · p. 39

9.5 Arabic Cuisine

  • 79.02% accuracy makes Jais 2 70B the top evaluated open model on Arabic cuisine, while Jais 2 8B scores 66.48% among models with no more than 13B parameters.The 70B model ranks first in four of six primary categories, including Geographic Provenance at 85.00%; it scores 50.00% on Religious Compatibility, trailing gemma-3-27b-it at 82.60%.Source: §9.5 · pp. 40–41, Table 17 · p. 40, Table 19 · p. 41

9.6 Arabic Poetry

  • Jais 2 70B has the highest average accuracy on Arabic poetry analysis and ranks first on 10 of 14 subtasks, while Jais 2 8B leads its size range and six subtasks.The 70B model exceeds Qwen2.5 72B and Gemma3 27B by more than 7 percentage points in average accuracy.Source: §9.6 · p. 42, Table 20 · p. 42

9.7 Islamic Question–Answering

  • Jais 2 70B scores 89.64% on PalmX and 89.10% on IslamicQA, achieving state-of-the-art performance on these culturally and jurisprudentially demanding benchmarks.Its results are more heterogeneous elsewhere: it scores 80.71% on QASI general Islamic knowledge and 81.38% on IslamicEval2025, while Qwen2.5-72B leads QASI inheritance; no training data from these benchmarks was used.Source: §9.7 · pp. 42–43, Table 21 · p. 44

9.8 Dream Interpretation

  • Jais 2 70B achieves 85.39% on Arabic-origin dream interpretation, while Jais 2 8B reaches 67.60% and ranks near the top of its size group.The benchmark measures culturally grounded symbolic interpretation; the reported results indicate strongest performance among large models for 70B and competitive performance within the mid-size group for 8B.Source: §9.8 · p. 46

9.9 Summarization

  • Jais 2 70B and 8B show strong summarization performance across Arabic and cross-lingual benchmarks, with 70B among the highest-performing open models on semantic fidelity.Evaluation uses ROUGE-LSum for content preservation and structural fidelity and BERTScore for semantic similarity; 8B matches or outperforms several models of similar or larger size.Source: §9.9 · p. 46, Table 23 · p. 47

9.10 Arabic Culture

  • Jais 2 70B has the highest average across four Arabic cultural benchmarks, ranks first on ArabicMMLU and ArabCulture, and Jais 2 8B is consistently near the top across them.Gemma-3-27B leads AraDice-Culture and Qwen2.5-72B leads DialectalArabicMMLU, where Jais 2 70B scores 67.32%; the average excludes Jawaher, which measures cultural and stylistic alignment.Source: §9.10 · p. 47, Table 24 · p. 48

9.11.1 Arabic IFEval Dataset

  • Arabic IFEval adapts English instruction-following tasks to Arabic and adds instructions for language-specific phenomena, with deterministic verification and strict and loose scoring.The dataset was manually reviewed and culturally adapted by Arabic linguists; evaluation reports prompt-level and instruction-level accuracy, and implicit requirements such as language appropriateness and coherence are foundational.Source: §9.11 · p. 48, §9.11.1 · p. 49, §9.11.2 · p. 49

9.11.3 Results and Analysis

  • Jais-2-70B scores 62.87% on Arabic prompt-level and 70.89% on instruction-level IFEval, compared with 52.97% and 62.50% for Jais-2-8B.Both models also perform strongly on English IFEval; the results indicate gains with scale on Arabic instruction following.Source: §9.11.3 · p. 49, Table 25 · p. 50

9.12.1 Small and Arabic-Centric Models (Jais 2-8B)

  • Jais 2 8B scores 72.48% on GSM8K and 80.82% on IFEval, outperforming or matching several models of similar or larger size on English tasks.It exceeds Aya-Expanse-8b on GSM8K (32.22%) and IFEval (46.04%), but trails some models on RACE and TruthfulQA.Source: §9.12.1 · p. 50, Table 26 · p. 51

9.12.2 Jais 2 vs Previous Jais Versions

  • Jais 2 8B improves over earlier Jais models by 20.85 points on GSM8K, 29.50 points on IFEval, and 7.17 points on MMLU, while trailing on DROP and RACE.The reported comparisons are GSM8K 72.48% vs 51.63%, IFEval 80.82% vs 51.32%, MMLU 62.95% vs 55.78%, DROP 10.53% vs 35.01%, and RACE 37.89% vs 44.98%.Source: §9.12.2 · p. 51

9.12.3 Large Models Comparison (Jais 2 70B)

  • Jais 2 70B leads evaluated models on HellaSwag (80.34%) and WinoGrande (75.85%), while remaining competitive across English benchmarks.The evaluations cover 12 English benchmarks in zero-shot settings. Jais 2 70B trails Llama-3.3-70B on GSM8K (85.97% vs. 93.78%) and Qwen2.5-32B on TruthfulQA (61.06% vs. 70.38%).Source: §9.12 · p. 50, §9.12.3 · pp. 51–52

10 Safety

Jais 2’s safety work combines curated Arabic offensive-language data, culturally targeted safety SFT, and safety prompting. The reported evaluations find that prompting reduces unsafe responses, especially for the 70B model.

10.1 Safety in Data Preparation

  • Thirty datasets were filtered into 205,125 samples, then organized under a unified taxonomy of non-offensive and offensive language.The taxonomy divides offensive content into general, obscene, and hate speech, with finer hate-speech categories; cleaning also removed duplicates, short samples, URLs, mentions, and other non-linguistic artifacts.Source: §10.1 · pp. 52–53
  • A quality-checked IFT dataset of 136,393 examples pairs templated instructions with explanations of why labeled content is offensive.More than 30 handcrafted templates and gold labels guided Modern Standard Arabic explanations; inconsistent samples were removed after a consistency check.Source: §10.1 · p. 53

10.2 Safety via SFT

  • Safety SFT data covers Arabic-relevant topics such as politics, religion, and economics alongside broader safety concerns.The work expanded a benchmark-inspired taxonomy from 13 risk types to 30 finer-grained types for broader, more precise coverage.Source: §10.2 · p. 54

10.2.1 General Risks

  • The five general-risk types were retained unchanged because they adequately covered the benchmark questions.They cover misinformation harms, human-chatbot interaction harms, malicious uses, discrimination/exclusion/toxicity/hateful or offensive content, and information hazards.Source: §10.2.1 · p. 55

10.2.2 Regional Risks

  • The eight regional risk types were refined into 25 subrisk types because the original categories were broader than the benchmark questions.The regional categories include cultural and social issues, governance, political dynamics, human rights, and controversial historical events.Source: §10.2.2 · p. 55

10.2.3 Prompt and Generation

  • Risk-specific and general instructions guided generation of 116,000 safety-focused QA pairs, split into about 50,000 general and 66,000 regional pairs.Questions could be challenging or accusatory, while responses were expected to remain safe and appropriate.Source: §10.2.3 · p. 55

10.2.4 Safety Evaluation Results

  • Human inspection found false positives in the original Arabic safety evaluation, prompting a regional-rubric revision to count unbiased, non-attacking responses as safe.The evaluation used around 5,800 Arabic Safeguard prompts and also assessed AraSafe and AraTrust; after revising Question 1, the authors used GPT-5 as the automatic judge.Source: §10.2.4 · pp. 57–59

10.3 Safety via Prompting

  • More than two-thirds fewer unsafe outputs were produced by Jais 2 70B with the safety prompt; prompting also improved both model sizes.The prompt was evaluated on the Arabic LLM Safeguard benchmark with Jais 2 8B and 70B, with and without the prompt.Source: §10.3 · p. 61, Table 29 · p. 62

11 Conclusion and Future Work

Jais 2 combines leading open-model results on key Arabic benchmarks with culturally grounded and general-purpose capabilities, strong instruction following, safety, and competitive English performance. The conclusion highlights open release and high-throughput deployment while identifying dialect coverage, reasoning, grounding, safety, and evaluation as future priorities.

Conclusion

  • Jais 2 achieved leading open-model results on OALL2 and AraGen while demonstrating cultural, general-purpose, instruction-following, safety, and English-language capabilities.The family includes a 70B model described as the largest open Arabic-centric LLM trained entirely from scratch to date and a highly competitive 8B variant.Source: §11 · p. 62
  • All models are openly released on HuggingFace under a commercially permissive license, and the 70B chat app reaches up to 2,000 tokens per second on Cerebras hardware.The application is available on the Web, iOS, and Android.Source: §11 · p. 62
  • Jais 2 is intended to support future Arabic-language research and deployment and contribute to a more inclusive multilingual AI ecosystem.The authors describe it as a foundation combining scale, linguistic diversity, cultural grounding, openness, safety alignment, and benchmark performance.Source: §11 · p. 62

Future Work

  • Future work will extend dialect and low-resource language support, improve long-context reasoning and factual grounding, strengthen culturally sensitive safety and alignment, and expand Arabic-centric benchmarks and datasets.The planned evaluation resources target cultural understanding, reasoning, and instruction following.Source: §11 · p. 62

12 Release Notes · Appendix · A List of Contributors

Jais 2 is released under Apache 2.0 for research and commercial use, with intended applications in Arabic-language research and products. The release notes also set limits around harmful or high-stakes use, language generalization, and potential bias.

12 Release Notes

  • Apache 2.0 enables research and commercial use, and the authors encourage experimentation to support reproducible research and Arabic language technology.They also aim to foster AI that better serves diverse linguistic and cultural communities.Source: §12 · p. 63

12.1 Intended Use

  • The model is intended for Arabic NLP research, businesses targeting Arabic-speaking audiences, and developers integrating Arabic capabilities into apps.The authors also identify academics researching Arabic NLP as a potential audience.Source: §12.1 · p. 63
  • Businesses can fine-tune Jais 2 for chat assistants, NLU/NLG tasks, customer service, and process automation.The passage names Jais 2 Chat as an example of a fine-tuned model.Source: §12.1 · p. 63

12.2 Out-of-Scope Use

  • Jais 2 should not be used to generate harmful content, promote hate or violence, spread misinformation, support illegal activity, or handle sensitive information.The prohibited scenarios include discrimination, fake news, and personal or confidential information.Source: §12.2 · p. 63
  • Because Jais 2 is optimized for Arabic and English, equal proficiency in other languages or dialects should not be assumed; high-stakes decisions require human oversight.The listed high-stakes areas include medical, legal, financial, and safety-critical decisions.Source: §12.2 · p. 63

12.3 Biases, Risks, and Limitations

  • Although Arabic data was curated and bias-reduction techniques were used, Jais 2 may still exhibit bias, so users should assess risks for their use cases before deployment.The model was trained on publicly available data.Source: §12.3 · p. 63

B Arabic Dream Interpretation Data Preparation · D) .ÈAÖÏ@ð QÓ B@ ú · ¯ H. @Q¢ “B@ð ¡J ªË@ úΫ ÈYK

The paper builds a structured Arabic dream-interpretation resource from classical and modern sources, then uses it to construct a culturally grounded multiple-choice benchmark with related alternatives. The benchmark includes Arabic examples and corresponding English translations.

B Arabic Dream Interpretation Data Preparation

  • The Arabic corpus contains 5,568 unique dream-interpretation pairs collected from classical references and online sources, retaining multiple interpretations when a symbol had them.The dataset covers a broad range of symbols and interpretations from Arabic dream-interpretation traditions.Source: B Arabic Dream Interpretation Data… · p. 65
  • Duplicate, incomplete, boilerplate, and cross-reference entries were removed, and the remaining pairs were normalized and manually reviewed for linguistic accuracy and consistency.Deduplication used exact and semantic matching, with one interpretation per line in the standardized format.Source: B Arabic Dream Interpretation Data… · p. 65
  • Symbols were assigned to 17 thematic categories through automatic classification followed by manual verification; Tools and Objects and Religious Symbols were the most frequent.The categories span domains including religious figures, natural elements, animals, food and drink, actions, emotions, and abstract concepts.Source: B Arabic Dream Interpretation Data… · p. 65
  • For each dream symbol, an Arabic natural-language question was generated with one correct interpretation and four distractors in structured Modern Standard Arabic JSON.The generation prompt is shown in Figure 16.Source: B Arabic Dream Interpretation Data… · p. 65, Figure 16 · p. 66
  • Distractors sampled from other interpretations in the same category keep choices thematically related and plausible, making the benchmark test discrimination among closely related alternatives.The benchmark is intended to assess symbolic understanding, culturally appropriate interpretation, and reasoning over contextually related Arabic alternatives.Source: B Arabic Dream Interpretation Data… · p. 65

¯ H. @Q¢ “B@ð ¡J ªË@ úΫ ÈYK

  • Representative benchmark examples pair the symbols Surat Al-Tin and Pain with Arabic questions and choices, alongside corresponding English translations.Each example includes the symbol, interpretation, generated question, answer options, and correct choice.Source: Figure 17 · p. 67, ¯ H. @Q¢ “B@ð ¡J ªË@ úΫ ÈYK · p. 67

C Arabic Cuisine Data Preparation

The section describes how Arabic recipes were collected, cleaned, and standardized into a curated dataset for instruction tuning and culinary evaluation. The workflow combines web crawling and preprocessing with model-assisted normalization and validation.

Data Collection

  • Five web sources were selected for coverage, cultural authenticity, and consistent formatting, then crawled to collect recipes reproducibly at scale.The sources represent diverse Arabic cuisines and regional cooking traditions.Source: C Arabic Cuisine Data Preparation · p. 68
  • Preprocessing removed irrelevant and duplicate entries, extracted recipe metadata, and normalized ingredient quantities, units, and formatting across sources.Optional fields such as preparation time, difficulty, yield, nutrition facts, images, and videos were preserved when available.Source: C Arabic Cuisine Data Preparation · p. 68
  • 64,987 recipes were curated across diverse dishes, ingredients, and cooking techniques for instruction tuning and Arabic culinary evaluation.The resulting dataset’s metadata coverage is summarized in Table 31, and a representative sample appears in Table 30.Source: C Arabic Cuisine Data Preparation · p. 68

Data Cleaning

  • Gemini Pro 2.5 converted recipe titles, ingredients, and steps into standardized JSON while cleaning text, resolving inconsistencies, and preserving culinary details.The cleaning workflow used two separate prompts, mergeable into one, as shown in Figure 18.Source: C Arabic Cuisine Data Preparation · p. 68, Figure 18 · p. 69

Validation and Error Handling

  • The cleaning prompt validated whether entries had clear ingredients and preparation steps, marking invalid entries false and recording a brief reason.Structured recipes support generating varied instruction-following examples and downstream tasks such as recipe understanding, procedural reasoning, and question answering.Source: C Arabic Cuisine Data Preparation · pp. 68, 70

D Arabic Cuisine Benchmark · 1. Question Generation · 2. Manual Validation

The Arabic Cuisine Benchmark comprises 1,803 recipe-based questions drawn from 601 curated canonical recipes, with question quality manually reviewed and strong overall annotator agreement. Its construction emphasizes broad culinary representation, semantic deduplication, and checks of question relevance, correctness, answer-option uniqueness, and subjectivity.

D Arabic Cuisine Benchmark

  • The benchmark contains 601 canonical recipes selected from roughly 4,000, yielding 1,803 multiple-choice items that test ingredients, preparation steps, and dish characteristics.Source: D Arabic Cuisine Benchmark · pp. 70–71
  • Curation sought coverage across Arab nations, religious dietary requirements, and varied lifestyles, balanced food categories, and canonical dishes rather than minor adaptations.For each country, the selection procedure identified five prominent recipes and then refined the set for diversity and fairness.Source: D Arabic Cuisine Benchmark · pp. 70–71
  • Canonical selection favored widely recognized standard recipes with essential identity-defining ingredients and used the most frequent median version when minor variants existed.Source: D Arabic Cuisine Benchmark · p. 71
  • Two-stage semantic deduplication used recipe embeddings and a 0.75 cosine-similarity threshold to cluster near-duplicates and retain one representative per cluster.Recipes without neighbors above 0.75 were retained as unique; embeddings were generated from titles, ingredients, and preparation steps, with FAISS used for nearest-neighbor search.Source: D Arabic Cuisine Benchmark · p. 71

1. Question Generation

  • Each recipe generated three self-contained MCQs, each with exactly five options: one correct answer and four plausible but incorrect distractors.Source: §1 · p. 71

2. Manual Validation

  • Three annotators manually reviewed every question for source relevance, gold-answer correctness, option uniqueness, and subjectivity, revising or discarding failures and resolving disagreements by majority vote.Source: §2 · p. 72
  • 0.81 overall Krippendorff’s alpha indicated strong reliability across annotators, with pairwise agreement of 0.77, 0.83, and 0.83.Source: §2 · p. 72
  • Agreement was perfect for relevance, moderate for subjectivity (α = 0.67), and lower for correctness (α = 0.49) and difficulty (α = 0.25).Source: §2 · p. 72

E Islamic QA Data Preparation

The Islamic QA data preparation pipeline uses Islamweb fatwas to create a cleaned instruction-tuning dataset and a manually reviewed 1,000-question multiple-choice benchmark spanning Islamic jurisprudence categories. Cleaning removes nonessential material while preserving rulings and scholarly evidence; referral screening filters out answers without standalone content.

Data Collection and Cleaning

  • Islamweb fatwas were scraped with question, answer, and available category metadata, then cleaned into concise, self-contained pairs without altering jurisprudential content or reasoning.The Gemini Flash 2.5 copy-editing prompt removed nonessential elements without summarizing or paraphrasing, while preserving Qur’an and Hadith evidence and scholars’ citations.Source: E Islamic QA Data Preparation · pp. 72–74

Data Cleaning

  • Referral screening filtered out answers without standalone content, leaving 151,890 samples for instruction fine-tuning data and benchmark creation.The screening checked whether each answer referred primarily to another fatwa or source rather than providing an independent ruling.Source: E Islamic QA Data Preparation · pp. 73–74

Benchmark creation

  • The 1,000-question benchmark covers 79 categories, with questions generated as Arabic MCQs, automatically checked for self-containment and ruling accuracy, and manually reviewed.The category selection considered importance in Islamic studies, complexity, and frequency; each MCQ has five plausible options and one correct answer.Source: E Islamic QA Data Preparation · p. 74

F Arabic Poetry Data Preparation … H Model Cards

This section documents Arabic poetry instruction-tuning data through dataset structure, task statistics, and instruction-format examples, and presents Arabic cuisine, safety, and Islamic Q&A examples alongside Jais 2 8B and 70B model-card tables.

F Arabic Poetry Data Preparation

G.1 Arabic Poetry Generation Examples

¯ iJ. K Xñƒ B@ I. ʾË@ àA¿ @ X@

G.3 Arabic Cuisine Examples

  • The appendix provides Arabic cuisine examples, though the supplied extracted passages do not support a reliable summary of their content.Source: §G.3 · p. 88

G.4 Safety Examples

  • For chemical or biological weapons and illegal weapons transfers, the examples refuse assistance and cite safety, illegality, or public-security concerns.Source: §G.4 · p. 89

G.5 Islamic Question Answering Examples

  • The appendix includes a table of general Islamic question-answering examples.Source: Table 38 · p. 91

H Model Cards

Related papers

Loading 2608.13580v1…