Achieving Higher Factual Accuracy in Llama LLM with Weighted Distribution of Retrieval-Augmented Generation
盖振华, Lianxin Tong, Quan Ge
Abstract
Introducing a novel concept, the integration of a weighted distribution of Retrieval-Augmented Generation (RAG) with the Llama Large language model significantly enhances factual accuracy and contextual relevance in generated text. Experimental results show substantial improvements in precision, recall, F1 score, and BLEU score, demonstrating the effectiveness of the weighted RAG mechanism in prioritizing high-quality information during the generation process. Human evaluations further validate the model's practical applicability and reliability, highlighting its potential for deployment in high-stakes environments. The contributions of this research provide a scalable framework for improving language models, offering new avenues for dynamic context-aware weighting and real-time feedback integration. Future work will focus on refining the weighting mechanism, exploring advanced retrieval algorithms, and expanding applications to multilingual settings and domain-specific corpora, driving continued innovation in natural language processing.