Jais: Arabic-Centric LLM Family

• Jais is an open-source, Arabic-centric LLM family optimized for Modern Standard Arabic, dialects, and English with models up to 30B parameters.
• It employs a GPT-3-style, decoder-only transformer architecture trained on 395B tokens using bilingual pretraining and instruction tuning.
• Jais demonstrates robust performance in dialect translation and cultural tasks, though challenges remain in safety alignment and handling low-resource dialects.

Jais is an open-source, Arabic-centric LLM family designed for deep natural language understanding and generation in Modern Standard Arabic (MSA), major Arabic dialects, and English. Jais models are decoder-only transformers with model sizes up to 13B and 30B parameters, engineered to address the scarcity of high-quality Arabic LLMs for tasks spanning machine translation, dialect handling, question answering, document classification, and stylometric robustness. Developed with a focus on both foundation pretraining and instruction tuning, Jais integrates advances in tokenization, bilingual representation learning, and safety alignment, and is actively benchmarked across emergent Arabic and cross-lingual tasks (Sengupta et al., 2023).

1. Model Architecture, Data, and Training Paradigm

Jais is a GPT-3-style, decoder-only transformer comprising 40 layers, 40 attention heads, and a model dimension $d_{\text{model}}=5\,120$, with $\approx$13B parameters. The feedforward layers use a depth of $d_{ff}\approx13\,653$, leveraging SwiGLU activations for computational efficiency. Positional information is encoded using ALiBi (Attention with LInear Biases), enhancing length generalization. The model is trained with pre-layer normalization, AdamW optimization, and a context window of up to 4,096 tokens (Sengupta et al., 2023, Al-Shaibani et al., 29 May 2025).

Jais’s training corpus totals 395B tokens and is composed of:

• 116B Arabic tokens (29%), with 55B native and 18B translated from English, and additional upsampling focused on Arabic data.
• 232B English tokens (59%), primarily from The Pile collection.
• 46B code tokens from GitHub repositories (12%).

The custom Jais tokenizer, a 84,992-subword BPE vocabulary, is trained on a 50/50 Arabic–English text split to optimize fertility for both languages and code (1.01–1.05) (Sengupta et al., 2023). The instruction-tuned variant (Jais-chat) is further fine-tuned on 10 million prompt–response pairs (6M English, 3.6M Arabic), using a mixture of widely used English datasets and numerous Arabic-specific sources.

2. Arabic, Dialectal, and Multilingual NLP Capabilities

Jais is specifically optimized for Arabic-first tasks but supports robust performance in English. Its instruction-tuned versions (Jais-chat) demonstrate pronounced gains in Arabic dialect identification, generation, and translation compared to multilingual and English-first LLMs:

• Dialect identification (micro-F1): Jais achieves 0.31 (QADI, 18-way), 0.40 (ADI, 14-way), and 0.58 (ADD, 4-way), outperforming multilingual LLaMA-3 and Mistral (0.26–0.42) and AceGPT in lower-resource settings, though lagging behind supervised SOTA by 20–30 F1 points.
• Dialect generation (accuracy): Jais attains 0.49–0.74, rivaling AceGPT in Gulf Arabic and substantially exceeding Llama 3/Mistral (0.32–0.63).
• Dialect→English MT (SacreBLEU): Jais yields 36–40 BLEU, matching or just below AceGPT and well above Llama 3 (12–27 BLEU); however, En→Dialect and MSA→Dialect remain challenging (<2.2 BLEU) (Mousi et al., 2024).
• Cultural MCQ tasks: Jais achieves ≈0.61 accuracy, indicating meaningful regional commonsense and domain knowledge above generic LLMs (Mousi et al., 2024).

Zero-shot evaluations on the ArabCulture commonsense benchmark reveal that base Jais models behave at random-guess level (~34% in MCQ), whereas Jais-chat yields 54–60% accuracy, outperformed by GPT-4o and Qwen-2.5-72B-Instruct (80–90%) (Sadallah et al., 18 Feb 2025).

3. General-Purpose NLP Benchmarks and Comparative Performance

Jais-13B-chat was tested on LAraBench across seven tasks (sarcasm detection, sentiment analysis, categorization, NLI, QA, subjectivity, gender identification) and consistently outperformed BLOOMZ on Arabic data (average F1=0.429 vs 0.391). While Jais slightly surpasses GPT-3.5-turbo in extractive QA (F1=0.546 vs 0.502), it trails both GPT-3.5 and GPT-4 (by 0.15–0.24 in F1) in classification tasks. Jais remains below the best fine-tuned task-specific Arabic models by 0.27 F1, indicating the persistent advantage of fully supervised approaches (Abdelali et al., 2023).

On cognitive and knowledge-intensive benchmarks, Jais obtains zero-shot accuracy of 46.5% (Arabic) and 53.9% (English), closely tracking open LLaMA2 and BLOOMz in English, but surpassing them on Arabic (Sengupta et al., 2023, Alshehhi et al., 25 Jul 2025). Compression experiments confirm that 4/8-bit quantization and moderate (≤40%) unstructured weight pruning preserve >96% of accuracy, with <2 percentage points performance drop and significant efficiency gains—demonstrating suitability for memory-, compute-, or deployment-constrained Arabic NLP (Alshehhi et al., 25 Jul 2025).

4. Machine Translation, Dialectal and Cultural Robustness

Jais demonstrates Arabic-centric advantages on dialect-sensitive MT and regional/cultural QA:

• In DA→MSA translation, Jais exhibits a lower need for post-editing than multilingual baselines (major-edit required in 30% of cases vs. 62% for NLLB-200), but struggles to fully resolve errors in dialectal terminology (TRM, ~29% of error mass) and semantic preservation (GSMIS, ~43%) (Alabdullah et al., 25 Dec 2025).
• Ara-HOPE analysis recommends augmentation of dialect-specific data, context-sensitive retrieval modules for DA expressions, and explicit semantic training objectives to address these bottlenecks (Alabdullah et al., 25 Dec 2025).
• In the AraDiCE benchmark, Jais outperforms all non-Arabic LLMs in dialect identification and knowledge tasks, but both Jais and AceGPT are constrained by training data sparsity in low-resource dialects (e.g., North Sudanese, SacreBLEU <2) (Mousi et al., 2024).
• On stylometric analysis, Jais-generated Arabic text displays consistent, model-specific lexical patterns (over-use of frequent tokens, shorter outputs), sustaining extremely high detection rates (>99% F1 in formal genres) unless prompted for informal, dialectal text (Al-Shaibani et al., 29 May 2025).

5. Safety, Alignment, and Limitations

Under standardized safety evaluation (SalamahBench), Jais 2 displays elevated vulnerability with a macro-averaged Attack Success Rate (ASR) of 25.8%, ranking last among five contemporary Arabic LLMs in exposure to unsafe generations. Category-level ASRs reach as high as 69.3% (intellectual property), and the lowest is 5.7% (privacy). Jais 2’s deficiencies are attributed to the absence of dedicated Arabic-first safety fine-tuning, limited exposure to negative examples, and lack of task-specific safety objectives. As a safety judge, Jais 2 attains only 17.5% accuracy (21/120), markedly below dedicated safeguard models. Recommendations include development of Arabic-native guardrails, expansion of dialectal and culturally nuanced harm data, and integration of policy-aware training objectives (Abdelnasser et al., 3 Feb 2026).

Systematic benchmarking confirms a broad pattern: Jais offers marked advances over non-Arabic-centric open models, particularly in dialect-handling and Arabic-specific tasks, but closed-source, larger, or more diversely supervised models retain a significant edge in cultural knowledge, controlled generation, and safety. For highly specialized or regulated domains (e.g., legal decision prediction), Jais-13B outperforms LLaMA-7B but remains half a point below GPT-3.5-turbo on expert-assigned human ratings (≈1.7/5 vs. ≈2.4/5), with a corresponding gap in BLEU/ROUGE metrics (Ammar et al., 2023).

6. Structured Prompting, Downstream Usage, and Best Practices

Instruction tuning (Jais-chat) substantially boosts performance on reasoning, cultural, and nuanced text generation. Chain-of-Thought (CoT) prompting elicits large F1 improvements on satirical news detection (up to 80% English F1, +26 F1 over zero-shot), with Jais-chat strongly benefiting from structured, stepwise inference—unlike multilingual models, which exhibit flat gains. The model excels most when prompt language matches target content and when CoT trades precision for high recall in ambiguous cases (Abdalla et al., 2024).

In practical terms, deploying Jais for Arabic NLP tasks should prioritize:

• Default 8-bit quantization for memory-limited settings.
• Moderate pruning for efficiency if inference speed is critical, while monitoring hallucination rates.
• Prompt engineering for task-specific performance (e.g., few-shot, CoT for complex classification).

This suggests that the core strengths of Jais derive from Arabic-centric pretraining, robust instruction tuning, and careful architectural balance, but its weaknesses in cultural commonsense, safety, and low-resource dialect generation persist without large, regionally diverse training data, richer negative examples, and targeted safety tuning.

7. Future Directions and Research Outlook

Current limitations in Jais highlight necessary future research: acquisition and integration of high-quality dialect-rich data across diverse Arabic communities; safety alignment specifically calibrated for Arabic harm categories; semantic and cultural augmentation via auxiliary objectives (e.g., contrastive learning on region- and culture-specific paraphrase pairs); and deeper studies of domain and style adaptation, especially for informal and low-resource genres.

Ongoing work also points to the importance of complementary safeguard systems, adversarial and red-teaming evaluation pipelines, and dynamic benchmarking that tracks model advances against emerging Arabic LLMs across all dialects, modalities, and application contexts (Sengupta et al., 2023, Abdelnasser et al., 3 Feb 2026, Sadallah et al., 18 Feb 2025).

---

References

(Sengupta et al., 2023, Mousi et al., 2024, Sadallah et al., 18 Feb 2025, Alabdullah et al., 25 Dec 2025, Abdelnasser et al., 3 Feb 2026, Abdelali et al., 2023, Alshehhi et al., 25 Jul 2025, Ammar et al., 2023, Al-Shaibani et al., 29 May 2025, Abdalla et al., 2024)