Fine-Tuning and Reasoning: Empowering LLMs for Solving Turkish Math Word Problems

Abstract

Large Language Models (LLMs) have gained significant attention due to their versatility across various tasks. This study investigates the adaptation of LLMs for solving Turkish mathematical word problems - a relatively underexplored area - through fine-tuning method. Three transformer-based models (Qwen 2.5, Mistral 7B, and Deepseek) were evaluated using a Turkish-translated version of the GSM8K dataset. Fine-tuning was performed using parameter-efficient techniques (LoRA), and models were assessed through zero-shot prompting with both quantitative metrics and qualitative analysis. Results show that Deepseek achieved the highest accuracy, while Qwen demonstrated strong step-by-step reasoning. The findings highlight the importance of language-specific fine-tuning, especially for morphologically rich languages like Turkish. This work lays a foundation for future research on multilingual LLM adaptation in mathematical reasoning tasks. © 2025 IEEE.