Boosting Domain Knowledge Understanding of LLMs through Fine-Tuning with a Novel KNN Algorithm

Abstract Transformer-based architectures have revolutionized natural language processing, demonstrating exceptional capabilities in generating and understanding human-like text. Despite these advancements, challenges persist in enabling models to effectively interpret and generate domain-specific knowledge with high accuracy. The novel integration of a k-nearest neighbour (KNN) algorithm with the GPT-Neo model represents a significant enhancement, specifically designed to improve the model's performance in specialized fields through capturing semantic similarities and contextual nuances inherent in domain-specific data. The research involved fine-tuning the GPT-Neo model with the KNN algorithm, resulting in substantial improvements in accuracy, coherence, and contextual understanding, as evidenced through higher F1, BLEU, and ROUGE scores, alongside reduced perplexity. Ablation studies highlighted the contributions of dynamic weighting schemes, custom distance metrics, and hybrid embeddings to the overall performance gains. The implications of this approach extend to various applications requiring precise domain knowledge interpretation, such as medical diagnostics, legal analysis, and technical support. Future work aims to explore broader domain datasets, integrate additional machine learning algorithms, and further refine the training methodologies to enhance model capabilities.