This thesis presents a comparative analysis of Llama 3-8B and DistilBERT language models for news classification across 26 classes. Utilizing a balanced dataset, we employed Low-Rank Adaptation (LoRA) for fine-tuning Llama 3-8B and traditional fine-tuning for DistilBERT. The study aims to evaluate the performance, efficiency, and practical applicability of these models in categorizing news articles.

Our experiments reveal that Llama 3-8B consistently outperforms DistilBERT in overall accuracy, achieving around 70% compared to DistilBERT's 60%. However, both models demonstrate competitive capabilities and exhibit distinct strengths across different news categories. The analysis uncovers significant variability in category-specific performance across multiple experimental runs, emphasizing the importance of robust evaluation procedures in model assessment.