Multilingual Student Performance Prediction Using Llama 3

Predicting student performance is a pivotal aspect of Educational Data Mining (EDM), facilitating early interventions and personalized learning strategies. Traditional models often struggle with multilingual and informal textual data, especially in regions with diverse linguistic backgrounds. This study introduces a novel approach that leverages the advanced capabilities of the Llama 3.1 model, a state-of-the-art large language model (LLM), to predict student grades in a multilingual context.We utilized a dataset comprising over 700 student records from various demographic regions across Sri Lanka, encompassing academic performance, personal ambitions, parental education levels, extracurricular activities, and learning styles. The data included textual entries in both Sinhala and English, often containing colloquial expressions. The Llama 3.1 8B model was employed to preprocess and translate this multilingual and informal textual data into coherent English, enhancing data consistency and quality.The model was fine-tuned using Transformer Reinforcement Learning (TRL) and Supervised Fine-Tuning (SFT) to predict grade classes (A, B, C, S, F) for the subsequent term. Our approach combines regression and classification methodologies and incorporates an enriched feature set, including learning style assessments and assistive learning participation. Hyperparameter tuning and k-fold cross-validation were applied to optimize model performance.Experimental results demonstrate that the fine-tuned Llama 3 model significantly outperforms traditional models like XGBoost and hybrid methods, achieving better results across multiple evaluation metrics and subjects. The model effectively handles the complexities of multilingual textual data and captures intricate feature interactions, leading to enhanced accuracy in student performance predictions.This study highlights the potential of advanced LLMs in EDM, particularly in multilingual settings, and suggests that integrating sophisticated language models can enhance the accuracy of student performance predictions. Future work will focus on extending this approach to other educational contexts and exploring the model’s ability to predict continuous scores for more precise assessments.