Classification of status of applicants based on other available datasets and parameters

Abstract

The study investigates the use of two well-known Large Language Model (LLM), Bidirectional Encoder Representations from Transformers (BERT) and LLaMA2, in evaluating student eligibility for the University of Stuttgart. By utilizing the capabilities of these models, our goal is to improve the efficiency, fairness, and scalability of student admission processes. BERT was chosen for its strong ability to comprehend contextual relationships in text using its bidirectional attention mechanism, enabling it to extract meaningful information from academic transcripts and categorize students based on their qualifications. LLaMA2, a generative model, was selected for its effectiveness in handling large-scale datasets through causal language modeling. The BERT model was fine-tuned for sequence classification by tokenizing academic records and inputting them into the model to determine a student’s qualification. Similarly, LLaMA2 was fine-tuned for causal language modeling using Low-Rank Adaptation (LoRA) and 4-bit quantization to reduce computational and memory overhead while maintaining high accuracy. The training process involved extensive hyperparameter tuning, including adjusting learning rates, batch sizes, and dropout rates, to achieve optimal performance. An 80-20 train-validation split was utilized to ensure that the models generalize well to new data. The main evaluation metric for BERT model was accuracy, precision, recall, and F1-score and for LLAMA2 it was a manual comparasion, providing a comprehensive assessment of the models’ performance across positive and negative classes. During model inference, the tokenized student data was passed through the trained models, which classified each student. The models also generated a probability score for each classification, enabling transparent and interpretable decision-making.