AraDiCE: Benchmarks for Dialectal and Cultural Capabilities in LLMs

TL;DR

AraDiCE addresses the scarcity of dialectal Arabic benchmarks by introducing a comprehensive, MT-plus-PEMT workflow to generate high-quality dialect datasets and a region-focused cultural benchmark across Gulf, Egyptian, and Levantine Arabic. The framework combines existing Arabic benchmarks translated into MSA and dialects with a novel AraDiCE-Culture dataset, enabling evaluation of understanding, generation, cognitive abilities, and culture-aware reasoning in LLMs. Key findings show Arabic-centric models outperform multilingual baselines on dialect tasks, yet dialect identification, generation, and translation remain challenging, and cultural nuance remains uneven across models. The study provides ≈45K post-edited samples, releases dialectal translation models, and highlights the need for dialect-tailored training to advance robust, culturally aware Arabic LLMs in real-world settings.

Abstract

Arabic, with its rich diversity of dialects, remains significantly underrepresented in Large Language Models, particularly in dialectal variations. We address this gap by introducing seven synthetic datasets in dialects alongside Modern Standard Arabic (MSA), created using Machine Translation (MT) combined with human post-editing. We present AraDiCE, a benchmark for Arabic Dialect and Cultural Evaluation. We evaluate LLMs on dialect comprehension and generation, focusing specifically on low-resource Arabic dialects. Additionally, we introduce the first-ever fine-grained benchmark designed to evaluate cultural awareness across the Gulf, Egypt, and Levant regions, providing a novel dimension to LLM evaluation. Our findings demonstrate that while Arabic-specific models like Jais and AceGPT outperform multilingual models on dialectal tasks, significant challenges persist in dialect identification, generation, and translation. This work contributes $\approx$45K post-edited samples, a cultural benchmark, and highlights the importance of tailored training to improve LLM performance in capturing the nuances of diverse Arabic dialects and cultural contexts. We have released the dialectal translation models and benchmarks developed in this study (https://huggingface.co/datasets/QCRI/AraDiCE).

Figures (15)

• Figure 1: Capabilities and associated datasets for benchmarking, evaluated on different dialects.
• Figure 2: Comparison on dialect identification
• Figure 3: Results on dialect generation
• Figure 4: Confusion matrices on dialect identification (QADI) dataset
• Figure 5: Average scores on ArabicMMLU