FinLLMs: A Framework for Financial Reasoning Dataset Generation with Large Language Models

• DOI: 10.48550/arxiv.2401.10744
• Publication Date: 2024-01-01
• Authors: Ziqiang Yuan; Kaiyuan Wang; Shoutai Zhu; Ye Yuan; Jingya Zhou; Yanlin Zhu; Wenqi Wei
• Keywords: Merge (version control); Benchmark (surveying)

Large Language models (LLMs) usually rely on extensive training datasets. In the financial domain, creating numerical reasoning datasets that include a mix of tables and long text often involves substantial manual annotation expenses. To address limited data resources reduce cost, we introduce FinLLMs, method for generating question-answering based common formulas using Models. First, compile list construct graph variables these employ. We then augment formula set by combining those share identical as new elements. Specifically, explore obtained merge with shared traversing constructed graph. Finally, utilizing GPT-3.5, generate encompasses both tabular information textual content, building collected set. Our experiments demonstrate synthetic generated FinLLMs effectively enhances performance several large-scale in outperforming two established benchmark