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 large-scale language modeling, while dialectal and culturally nuanced contexts pose additional challenges. Jais 2 develops Arabic-centric models and reports leading open-model results on key Arabic benchmarks, alongside open release and high-throughput deployment.

  • Problem

    Arabic is underrepresented in large-scale language modeling, and models trained mainly on MSA and English fail to capture users’ local dialects.

  • Method

    Jais 2 develops an Arabic-centric model family including 70B- and 8B-parameter variants, with domain upsampling used as a late-training data intervention.

  • Results

    Jais 2 achieves leading results among evaluated open models on OALL2 and AraGen, with strong performance on culturally grounded Arabic tasks and general tasks.

  • Takeaways & Limitations

    Open release under a commercially permissive license and deployment of Jais 2 70B at up to 2,000 tokens per second support accessible Arabic AI development and serving.

  • Takeaways & Limitations

    Despite bias-reduction efforts, the model may still exhibit biases, so users should assess relevant biases before deployment.

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 advances Arabic-centric language modeling through scale, Arabic-focused data, architectural and training improvements, cultural grounding, and open deployment. It achieves strong Arabic and competitive English performance while providing publicly accessible models and applications.

  • Data and Language Coverage: The models combine domain-diverse Arabic data, dialectal and script coverage, Arabizi, and bilingual Arabic–English training.The training emphasizes culturally grounded Arabic domains while retaining competitive English performance.
  • Architecture and Training: Jais 2 uses an Arabic-centric vocabulary, expanded feedforward layers, ReLU2 activations, µP, and multi-stage post-training with DPO and GRPO.These choices target efficient scaling, instruction following, and safety alignment.
  • Models: Jais 2 releases 8B and 70B open-weight Arabic-centric models trained entirely from scratch.The 70B model is described as among the largest open Arabic-centric models trained from scratch, while the 8B model is competitive among similarly sized open models.
  • Evaluation: Jais 2 obtains leading or competitive results on OALL2, AraGen, and multiple culturally grounded Arabic tasks, while remaining competitive in English.Reported domains include poetry, cuisine, dream interpretation, translation, and summarization.
  • Openness and Deployment: Both model sizes are released on Hugging Face under a commercially permissive license, with web and mobile access also provided.The 70B chat application is available on the Web, iOS, and Android and runs on Cerebras hardware at up to 2,000 tokens per second.
  • Conclusion: Jais 2 is positioned as an open-weight foundation for Arabic-centric research and development.The stated emphasis combines scale, linguistic diversity, cultural grounding, openness, and deployment efficiency.

2 Related Work

Related work frames Arabic-centric modeling as a response to multilingual systems’ English bias and limited handling of Arabic morphology, diglossia, dialects, and cultural context. Prior efforts span bilingual models, dialect-specific systems, instruction-tuning resources, and multidimensional Arabic benchmarks.

  • Arabic Language Models: Multilingual LLMs often underperform on Arabic because of English bias and Arabic’s complex morphology and diglossia.General-purpose models are described as especially limited in dialectal and culturally nuanced contexts.
  • Arabic Language Models: Arabic-centric models have explored bilingual pretraining, cultural alignment, instruction following, and regional adaptation.Examples include Jais, Jais-Chat, AceGPT, ALLaM, and Fanar.
  • Arabic Training Data: Arabic instruction-tuning resources combine native Arabic supervision, translated English instructions, and task mixtures for general utility.Related work also includes multilingual instruction corpora such as Aya, xP3, and Tülu-style datasets.
  • Arabic Training Data: Jais 2 builds on prior work with bilingual instruction tuning anchored in Arabic data spanning MSA, 17 dialects, Arabizi, poetry, religion, and science.This approach explicitly targets dialectal diversity and script variation.
  • Arabic Evaluation Benchmarks: Arabic evaluation has converged on knowledge and reasoning, instruction following and generation, and safety and bias.These axes use multiple-choice benchmarks, open-ended judging protocols, and culturally aware audits.

3 Model

Jais 2 uses a decoder-only Transformer with Arabic-oriented vocabulary and architectural choices selected through scaling experiments and ablations. Its training setup combines µP-based scaling, AdamW optimization, long-context and data-processing choices, and Cerebras infrastructure.

  • Architecture: Jais 2 uses a standard decoder-only Transformer with width, depth, and intermediate-size choices guided by empirical ablations and scaling laws.The width-to-depth ratio is kept near 100, while the feedforward filter size is set to eight times the hidden size.
  • Architecture: The model adopts decoupled input and output embeddings, RoPE, and a longer training context than Jais 1.Controlled ablations found RoPE superior to ALiBi at training length and longer zero-shot contexts.
  • Scaling and Optimization: µP enables zero-shot transfer of optimal hyperparameters from small proxy models to larger models.The search covers learning rates, initialization scales, embedding and unembedding scalars, and per-layer-type parameters.
  • Training Infrastructure: Training and post-training experiments ran on Condor Galaxy clusters containing 64 interconnected Cerebras CS-2 systems each.The infrastructure includes MemoryX for model states and SwarmX for broadcasting weights and aggregating gradients.
  • Tokenizer: The tokenizer has a 150,272-token vocabulary built with HuggingFace tokenizers and sampled primarily from English and Arabic data.Additional languages and programming text support broader coverage while preserving tokenizer compactness and generalization.
  • Tokenizer: Pre-tokenization combines a designed regular expression with ByteLevel processing to preserve byte alignment and space-prefixed tokens.The design segments contractions, alphanumerics, punctuation, and long whitespace spans.

4 Pretraining

Jais 2 pretraining combines staged learning, Arabic-focused data curation, multilingual corpus design, and domain upsampling. The resulting data and training interventions target Arabic coverage, cross-lingual transfer, and focused capability improvement.

  • 4 Pretraining: Jais 2 uses two-stage pretraining, beginning with diverse web data and followed by targeted, high-quality domain data.Stage 1 typically consumes most training FLOPs, while Stage 2 uses a smaller budget to address weaknesses identified earlier.
  • 4 Pretraining: The updated Arabic preprocessing pipeline normalizes documents, retains those with non-majority noise, cleans document content, and removes empty or overly short outputs.Changes from Jais 1 include explicit textual conversion of encoded religious-expression symbols and removal of JavaScript fragments and very long URLs.
  • 4 Pretraining: Arabic data spans 17 dialectal variants and includes supervised subsets for poetry, literature, religion, and Arabic scientific content.The dialectal content totals approximately 600 million tokens and is dominated by Moroccan, Egyptian, Gulf, Iraqi, and Emirati Arabic.
  • 4 Pretraining: 624B Arabic tokens combine a reprocessed Jais 1 corpus with specialized datasets, including a 26B-token high-quality religious subset.The religious data received expert-driven manual review and refinement for accuracy and cultural sensitivity.
  • 4 Pretraining: At 6.7B scale, an Arabic:English ratio of 1:1.5 improved Arabic accuracy when all Arabic tokens were used, leading to a final Arabic fraction of 28%.Fixed-token ablations identified 1:2 as a strong trade-off, while the adjusted ratio supported full Arabic-token consumption and cross-lingual transfer.
  • 4 Pretraining: Annealing-phase upsampling consistently improved Arabic, English, and math scores, producing 5–9% average gains across model sizes.The strongest gains occurred in the specifically upsampled domains, and US Mix 4 improved English and Arabic without significant math regression.

5 Post-Training

Jais 2 post-training refines domain expertise, instruction following, and preference alignment through continual pretraining, instruction fine-tuning, and DPO. These stages combine curated Arabic resources, replay, synthetic bilingual data, and culturally rich tasks.

  • 5 Post-Training: Post-training consists of continual pretraining, instruction fine-tuning, and preference alignment.The sequence is designed to refine pretrained capabilities and alignment.
  • 5 Post-Training: Continual pretraining uses curated Arabic data and replay from the original corpus to improve target-domain expertise and weaker capabilities.The stage runs for two epochs using a mixture of newly constructed data and original pretraining data.
  • 5 Post-Training: Instruction fine-tuning uses over 20 million instruction–completion pairs spanning Arabic and English resources, synthetic data, and culturally rich tasks.Examples include dream interpretation and Arabic poetry; high-quality English datasets were translated into Arabic when needed.
  • 5 Post-Training: Direct Preference Optimization aligns model behavior with human preferences and ethical principles for safer, more helpful assistance.The reported post-training combination provides linguistic grounding, instruction following, and alignment with human values and cultural context.

6 Continual Pretraining

Continual pretraining strengthens Jais 2 in specialized domains and addresses weaknesses identified after initial instruction fine-tuning. It combines curated Arabic data with replay and broad synthetic bilingual content.

  • 6 Continual Pretraining: Continual pretraining runs for two epochs on a newly constructed corpus with 50% replay from the original pretraining dataset.Replay preserves exposure to the original distribution while the new corpus supports further specialization.
  • 6 Continual Pretraining: The stage targets both frontier-level performance in specialized domains and weaker competencies revealed by initial instruction fine-tuning.Its objectives explicitly separate specialization from remediation of post-IFT weaknesses.
  • 6 Continual Pretraining: Specialized curated Arabic datasets support domain expertise, while synthetic English and Arabic data broadens coverage across domains and topics.Synthetic resources include textbook-style content, explanation-enriched multiple-choice questions, and mathematics datasets.

7 Instruction Fine-Tuning

Jais 2’s instruction fine-tuning combines broad bilingual supervision with targeted dialectal and culturally grounded datasets. The resulting training design supports instruction following across Arabic varieties, cultural domains, and tasks including translation, poetry, cuisine, and dream interpretation.

  • 7 Instruction Fine-Tuning: IFT addresses the gap between next-token pretraining and following natural-language instructions through supervised prompt–response training.The paper also calls IFT supervised fine-tuning (SFT).
  • 7.1 General IFT: Standard Arabic and English: The general IFT mixture contains over 20M diverse data points spanning English, standard and dialectal Arabic, synthetic capabilities, and Arabic culture-centric collections.The sources include rewritten Jais 1 SFT data, public and synthetic datasets, and culturally focused Arabic material.
  • 7.1 General IFT: Standard Arabic and English: Synthetic training data covers multi-turn dialogue, safety, logic, mathematics, sciences, translation, Arabic grammar, and sentiment analysis.Multi-turn data targets reference resolution, recap ability, context retention, and knowledge adaptation.
  • 7.1 General IFT: Standard Arabic and English: The generated training data was decontaminated against benchmark test examples, and additional system-prompt examples taught consistent adherence to system prompts.The decontamination procedure evaluated each synthetic training example against benchmark test data before removing contaminated examples.
  • 7.2 General IFT: Dialectal Arabic: Dialectal IFT targets Darija and Egyptian Arabic through systematic collection, annotation, and validation of linguistically representative data.The datasets capture distinctive lexical, morphological, and syntactic features of the two dialects.
  • 7.3 Task: Dialectal Arabic Translation: Dialectal translation data bridges MSA–English resources and lower-resource colloquial Arabic across diverse MENA varieties, improving query understanding and response accuracy.The corpus is designed to handle local linguistic nuances across the Middle East and North Africa.
  • 7.4–7.8 Culturally Grounded Tasks: The curated instruction-tuning program covers dialect identification, cuisine, poetry, and dream interpretation through task-specific datasets and culturally grounded benchmarks.Dialect identification uses 624K examples covering 15 dialects and languages; cuisine evaluation uses recipe-grounded multiple-choice reasoning.

8 Preference Alignment

Jais 2 uses Direct Preference Optimization to align model behavior with human preferences without an external reward model. The curated preference data and tuned training configuration improved English and Arabic instruction following and Vicuna evaluation win rates.

  • DPO Method: Direct Preference Optimization aligns language models directly from preference data without relying on an external reward model.DPO is described as a preference-alignment stage for Jais 2.
  • DPO Method: DPO compares preferred and rejected responses using log-likelihood ratios from the optimized policy and a fixed reference policy, typically the SFT model.The Bradley–Terry formulation determines preference for a response given a prompt.
  • DPO Objective: The DPO objective increases the relative likelihood of preferred responses while implicitly regularizing the optimized policy toward the reference policy.The temperature β controls update strength and σ denotes the sigmoid function.
  • Preference Data: The preference dataset contains over 200k chosen-and-rejected pairs spanning general conversation, Arabic, mathematics, and instruction following.The data was expanded through regenerated instructions, self-play responses, and frontier-LLM judging.
  • Results: DPO improved English and Arabic instruction-following performance and increased win rate in Vicuna evaluations.The passage reports this as the observed outcome of the DPO stage.

9 Evaluation

Jais 2 performs strongly across Arabic evaluation settings, including standardized knowledge tasks, translation, dialect identification, and culturally grounded benchmarks. The 8B model is competitive within its size range, while the 70B model frequently leads evaluated open models.

  • OALL2: 72.40% macro-average makes Jais-2-8B the highest-scoring model among those with at most 13B parameters on OALL2.It outperforms Fanar-1-9B-Instruct at 68.97% and ALLaM-7B-Instruct-preview at 67.29%.
  • OALL2: 79.36% overall average makes Jais-2-70B the best model above 13B parameters on OALL2.It exceeds Llama-3.3-70B-Instruct at 74.23% and Qwen2.5-72B-Instruct at 71.20%, with highest accuracy on AlGhafa, ArabicMMLU, AraTrust, and MadinahQA.
  • Arabic Translation: Jais 2 70B achieves the strongest open-model translation performance across MSA, English, and multiple Arabic dialect directions.Jais 2 8B is among the strongest models below 13B, while the 70B model consistently reaches the highest or near-highest BLEU scores across dialect pairs.
  • Arabic Dialect Identification: Jais 2 70B leads Arabic dialect identification on MADAR and QADI, while Jais 2 8B records 46.03% and 44.24% average accuracy, respectively.The 70B model ranks first across nearly all dialect categories; the 8B model remains competitive within its parameter group.
  • Culturally Grounded Evaluation: Jais 2 70B reaches 79.02% on Arabic cuisine classification and leads 10 of 14 subtasks in a culturally grounded evaluation.It ranks first in four cuisine categories but scores 50.00% on Religious Compatibility, trailing gemma-3-27b-it at 82.60%.

10 Safety

Jais 2 strengthens Arabic safety through curated offensive-language data, unified taxonomies, explanatory instruction-following examples, and culturally grounded safety training across regional and general risks.

  • Accurately identifying offensive language supports safer, more reliable Arabic-model interactions across dialects and sensitive domains.
  • Offensive Language Data: Thirty publicly available datasets were manually inspected, with unreliable sources excluded before aggregation into the offensive-language corpus.
  • Taxonomy and Cleaning: A unified hierarchical taxonomy merges overlapping labels into non-offensive, general, obscene, and hate-speech classes with finer-grained hate categories.
  • Taxonomy and Cleaning: Deduplication, minimum-length filtering, and removal of non-linguistic artifacts reduce redundancy and improve the quality of training examples.
  • Contextual Enrichment and IFT: Explanatory responses replace simple offensive/non-offensive labels, enabling more informative and context-aware model behavior.
  • Safety Alignment: The safety pipeline expands risk coverage to 30 finer-grained types spanning Arabic-relevant regional concerns and broader safety topics.
  • Safety Evaluation: Human evaluation found that prior Arabic-versus-English vulnerability results were affected by a high rate of false positives in safety judgments.

11 Conclusion and Future Work

Jais 2 combines strong Arabic-centric and multilingual capabilities with open access and high-throughput deployment. The authors identify broader dialect coverage, stronger reasoning and factual grounding, and expanded safety evaluation as future priorities.

  • Conclusion: Jais 2 achieves leading Arabic benchmark results while maintaining strong culturally grounded, instruction-following, safety, and English-language capabilities.
  • Openness and Deployment: All model weights are openly released under a commercially permissive license, while the 70B model is also available through web and mobile chat applications.
  • Openness and Deployment: Cerebras deployment enables inference speeds of up to 2,000 tokens per second for the 70B chat application.
  • Conclusion: The models provide a foundation for future Arabic language technologies, research, practical deployment, and broader Arabic-focused innovation.
  • Future Work: Future work includes additional dialects and low-resource varieties, longer-context reasoning, factual grounding, stronger safety alignment, and broader Arabic-centric benchmarks.

12 Release Notes

Jais 2 models are released openly for research and commercial use, with users expected to follow licensing, legal, and organizational requirements. The release notes also define important boundaries around misuse, language coverage, high-stakes decisions, and residual bias.

  • The Jais 2 models are released under the Apache 2.0 license for research and commercial use.
  • Users are encouraged to experiment with, evaluate, and build upon the models while complying with applicable laws, regulations, policies, and license terms.
  • Out-of-Scope Use: Jais 2 should not be used to generate harmful content, handle sensitive information, or support illegal, discriminatory, or misleading activities.
  • Out-of-Scope Use: Jais 2 is optimized for Arabic and English, so equal proficiency in other languages or dialects should not be assumed.
  • Out-of-Scope Use: High-stakes medical, legal, financial, or safety-critical decisions require human oversight rather than unaided model use.
  • Limitations: Although preprocessing sought to reduce bias, residual biases may remain because the model was trained partly on publicly available Arabic data.

B Arabic Dream Interpretation Data Preparation

The Arabic dream interpretation resource was built by collecting, cleaning, categorizing, and structuring interpretations from classical and modern sources. Its Arabic MCQ benchmark uses controlled generation with plausible same-category distractors to evaluate culturally grounded symbolic understanding.

  • The dataset combines classical Islamic and modern online sources through collection, cleaning, consolidation, categorization, and task formulation.
  • Data Sources: Five widely cited classical references and two primary websites supplied dream symbols and interpretations, with multiple interpretations retained when available.
  • Cleaning and Consolidation: Duplicate and incomplete records were removed, boilerplate was deleted, entries were normalized, and manual review checked linguistic accuracy and consistency.
  • Symbol Categorization: Each symbol was assigned to one of 17 thematic categories using a controlled prompt followed by manual verification.
  • MCQ Benchmark Construction: The benchmark automatically generates Arabic questions with one correct interpretation and four distractors for each dream symbol.
  • MCQ Benchmark Construction: Distractors come from other interpretations in the same category, making options thematically related and semantically plausible.
  • The benchmark assesses symbolic meaning, culturally appropriate interpretation selection, and reasoning among contextually related Arabic alternatives.

C Arabic Cuisine Data Preparation

The Arabic cuisine dataset was collected from multiple web sources and transformed into structured recipe records through cleaning, normalization, extraction, and validation. The resulting collection contains 64,987 recipes for instruction tuning and culinary evaluation.

  • Data Collection: Five web sources and automated crawlers provided broad, reproducible coverage of Arabic cuisines and regional cooking traditions.
  • Data Cleaning: The preprocessing pipeline filtered irrelevant pages and duplicates while extracting recipe titles, ingredients, cooking steps, and optional metadata.
  • 64,987 recipes remained after cleaning, spanning diverse dishes, ingredients, and cooking techniques.
  • Cleaning Prompt: The workflow uses expert-prompted JSON generation to clean, normalize, and extract recipe information from titles, ingredients, and preparation steps.
  • Structured Field Extraction: Structured fields include recipe names, numbered preparation steps, step summaries, ingredient text, ingredient quantities, and misplaced steps.
  • Validation and Error Handling: Validation marks entries without clear ingredients or preparation steps as invalid and records a reason for rejection.

D Arabic Cuisine Benchmark

The Arabic cuisine benchmark selects diverse canonical recipes and generates three self-contained MCQs per recipe. Manual review supports strong overall annotation reliability, but agreement is lower for correctness and difficulty.

  • Benchmark Construction: 601 canonical Arabic recipes were selected from roughly 4,000 candidates, producing 1,803 evaluation items with three questions per recipe.
  • Selection Criteria: The benchmark targets culinary diversity across Arab nations, food categories, religious dietary requirements, and selected nutritional needs.
  • Canonical Fairness: Canonical selection prioritizes culturally standard recipes, essential ingredients, and median versions over minor regional variations.
  • Recipe Selection: Semantic deduplication uses recipe embeddings and FAISS similarity search with a cosine similarity threshold of 0.75.
  • Question Generation: Each question is self-contained and contains exactly five options: one correct answer and four contextually relevant distractors.
  • Benchmark Validation: Three annotators manually reviewed every question for relevance, correctness, option uniqueness, and subjectivity, revising or discarding failed items.
  • Benchmark Validation: 0.81 Krippendorff’s alpha indicates strong overall annotation reliability, while correctness and difficulty agreement were lower at α = 0.49 and α = 0.25.

E Islamic QA Data Preparation

The Islamic QA data was scraped, cleaned into self-contained question–answer units, and filtered to remove referral-only answers. A 1,000-question MCQ benchmark was then generated, validated, and manually reviewed for evaluation in Islamic jurisprudence.

  • Data Cleaning: Cleaning removed boilerplate, personal elements, external references, and formatting noise while preserving scholarly evidence and original meaning.The process retained Qur’an, Hadith, and scholarly citations when they were part of the source content.
  • Data Cleaning: 151,890 samples remained after referral screening for IFT data and benchmark creation.
  • Benchmark Creation: The benchmark targeted comprehension and reasoning through multiple-choice questions covering the sensitive and diverse domain of Islamic jurisprudence.Prompt-driven validation was complemented by manual verification.
  • Benchmark Creation: 1,000 question–answer pairs were selected across 79 categories, emphasizing 20 major Islamic knowledge categories.
  • MCQ Generation and Validation: Each generated Arabic MCQ used five plausible options with exactly one correct answer, followed by automated checks and revisions when needed.Validation tested whether questions were self-contained and whether answers matched the underlying Islamic ruling.
  • MCQ Generation and Validation: All 1,000 benchmark samples underwent manual review for accuracy and appropriateness across the religious topics covered.The review served as the final quality-assurance step.

F Arabic Poetry Data Preparation

The Arabic poetry IFT dataset organizes poetic understanding and composition into analysis and generation tasks, with detailed subtasks and varied metadata-based instruction formats. Its tables document dataset statistics and evaluation coverage across these task configurations.

  • Task Organization: The Arabic poetry IFT dataset spans two core tasks: Analysis and Generation.The tasks are designed to capture different dimensions of poetic understanding and composition.
  • Task Organization: Detailed subtasks expose diverse input–output configurations involving metadata, linguistic features, and stylistic dimensions.
  • Task Organization: Instruction templates incorporate poem text, poet name, meter, rhyme, and genre as relevant task information.
  • Evaluation: Table 20 provides complete evaluation results for all Analysis subtasks and their range of prediction targets.
  • Dataset Statistics: Table 34 reports overall poetry IFT statistics across tasks and data splits.
  • Dataset Statistics: Table 35 reports combined statistics for analysis, continuation, restoration, and generation subtasks.
Loading 2608.13580v1…