The use of different prompt formats to improve the performance of language models in the turkish language

Language models can be trained and customized by anyone using effective finetuning methods, making them versatile tools across various domains. While metrics like loss and accuracy are commonly used to assess the performance of language models, the choice of prompt formats also plays a crucial role. In this study, large language models such as Llama2, Llama3 and Mistral7B, as well as the smaller Phi3 model, were trained and evaluated on two different datasets using both commonly used prompt formats and customized Turkish prompt formats. Based on the findings and results obtained, it has been demonstrated that not only the model architecture but also the structure of the prompt format plays a critical role in performance, and that changes in prompt formats can affect the performance of language models. Furthermore, the customized prompt format achieved the best metric results within acceptable loss values.