Science Exam Question Answering based on Retrieval-Augmented Generation

With the increasing capabilities of large language models (LLMs), researchers are exploring novel ways to leverage them in various domains. This competition explores the potential of large language models in answering challenging science-based questions. In response to the inherent limitations of scant training data, our proposed solution introduces a pioneering Retrieval-Augmented Generation (RAG) approach, incorporating the expansive Wikipedia6.5M dataset. Our innovation centers on skillfully incorporating external knowledge, a crucial factor in improving question-answering in limited computational settings. The synergy of vector similarity retrieval and Platypus2-70B LLM broadens perspectives, enhancing comprehension of STEM subjects and overcoming data scarcity by utilizing external knowledge repositories.