Request Correction

Instruction-Tuned Urdu LLMs: Efficient Adaptation of Llama Models and Evaluation Resources for Urdu

This paper presents UrduLLaMA 1.1 and UrduLLaMA 1.1 Tiny, two instruction-tuned large language models (LLMs) designed to advance natural language processing for Urdu, a low-resource language with limited representation in multilingual corpora. These instruction-tuned models are derived from Llama-3.1-8B-Instruct and Llama-3.2-3B-Instruct architectures, respectively by conducting continual pretraining on 800 million diverse Urdu tokens curated from public and proprietary sources, followed by Supervised Fine-Tuning (SFT) using LoRA on 432K Urdu instructions spanning diverse NLP tasks. Rigorous evaluation across 14 culturally-specific domains using our novel Urdu LLM Evaluation Dataset demonstrates superior performance. UrduLLaMA 1.1 achieves 65.3 average accuracy (GPT-5 Nano evaluation), outperforming its Llama-3.1-8B-Instruct base (50.7) across all categories and surpassing Llama-3.3-70B-Instruct (62.7) in 8 out of 14 domains. UrduLLaMA 1.1 Tiny transforms Llama-3.2-3B-Instruct (38.8) into a (61.2) performer. Human evaluation by native Urdu linguists confirms these gains (3.51/5 vs. 2.61/5 base). Our results validate targeted adaptation strategies combining continual pretraining with instruction tuning as computationally efficient solutions for low-resource languages, enabling state-of-the-art Urdu LLM models with accessible hardware.