ScholarGPT: Fine-tuning Large Language Models for Discipline-Specific Academic Paper Writing

Chuxue Cao, Ziqing Yuan, Hailiang Chen

Abstract: Description Despite the impressive capabilities demonstrated by generative AI tools like ChatGPT in content generation, they exhibit certain limitations when it comes to assisting business research writing. These limitations encompass factual errors and instances of hallucination, which can be attributed to insufficient domain-specific data during the pre-training stage. To address these challenges, we have constructed a comprehensive Business Research Instruction-following dataset (BRI) comprising 158,254 prompt-completion pairs extracted from academic papers in the business field. Building upon this dataset, we have developed ScholarGPT, a discipline-specific academic writing assistant that leverages fine-tuned large language models. ScholarGPT aims to aid researchers in their academic writing endeavors, specifically during the initial drafting stage. It is designed to propose relevant titles, keywords, abstracts, outlines, hypotheses and hypothesis development based on the inputs provided. Through our experimental evaluations, we find that ScholarGPT surpasses Llama-2-13b-chat and gpt-3.5-turbo in various domain-specific academic writing tasks.