UrduBench: An Urdu Reasoning Benchmark using Contextually Ensembled Translations with Human-in-the-Loop

Recent advances in large language models (LLMs) have led to strong reasoning capabilities; however, evaluating such models in low-resource languages remains challenging due to the lack of standardized benchmarks. In particular, Urdu reasoning evaluation has been limited by the sensitivity of machine translation and an emphasis on general language tasks rather than reasoning benchmarks. In this paper, we propose a contextually ensembled translation framework with human-in-the-loop validation that leverages multiple translation systems to develop Urdu reasoning benchmarks while preserving contextual and structural integrity. Using this framework, we translate widely adopted reasoning and question-answering benchmarks, including MGSM, MATH-500, CommonSenseQA, and OpenBookQA, into Urdu, collectively referred to as UrduBench, and conduct a comprehensive evaluation of both reasoning-oriented and instruction-tuned LLMs across multiple prompting strategies. Our analysis reveals performance differences across (1) four datasets, (2) five task difficulty levels, (3) diverse model architectures, (4) multiple model scaling settings, and (5) language consistency tests. We find that multi-step and symbolic reasoning tasks pose significant challenges in Urdu, and that stable language alignment is a critical prerequisite for robust reasoning. Overall, our work establishes a scalable methodology for standardized reasoning evaluation in Urdu and provides empirical insights into multilingual reasoning failures. This experimental setup is also broadly applicable to other low-resource languages. The code and datasets will be publicly released.

UrduLLaMA 1.0: Dataset Curation, Preprocessing, and Evaluation in Low-Resource Settings

Multilingual Large Language Models (LLMs) often provide suboptimal performance on low-resource languages like Urdu. This paper introduces UrduLLaMA 1.0, a model derived from the open-source Llama-3.1-8B-Instruct architecture and continually pre-trained on 128 million Urdu tokens, capturing the rich diversity of the language. To enhance instruction-following and translation capabilities, we leverage Low-Rank Adaptation (LoRA) to fine tune the model on 41,000 Urdu instructions and approximately 50,000 English-Urdu translation pairs. Evaluation across three machine translation datasets demonstrates significant performance improvements compared to state-of-the-art (SOTA) models, establishing a new benchmark for Urdu LLMs. These findings underscore the potential of targeted adaptation strategies with limited data and computational resources to address the unique challenges of low-resource languages.

Benchmarking Pre-trained Large Language Models' Potential Across Urdu NLP tasks

Large Language Models (LLMs) pre-trained on multilingual data have revolutionized natural language processing research, by transitioning from languages and task specific model pipelines to a single model adapted on a variety of tasks. However majority of existing multilingual NLP benchmarks for LLMs provide evaluation data in only few languages with little linguistic diversity. In addition these benchmarks lack quality assessment against the respective state-of the art models. This study presents an in-depth examination of prominent LLMs; GPT-3.5-turbo, Llama2-7B-Chat, Bloomz 7B1 and Bloomz 3B, across 14 tasks using 15 Urdu datasets, in a zero-shot setting, and their performance against state-of-the-art (SOTA) models, has been compared and analysed. Our experiments show that SOTA models surpass all the encoder-decoder pre-trained language models in all Urdu NLP tasks with zero-shot learning. Our results further show that LLMs with fewer parameters, but more language specific data in the base model perform better than larger computational models, but low language data.