Data-Centric Financial Large Language Models

• DOI: 10.48550/arxiv.2310.17784
• Publication Date: 2023-01-01
• Authors: Zhixuan Chu; Huaiyu Guo; Xinyuan Zhou; Yijia Wang; Fei Yu; Hong Chen; Wanqing Xu; Xin Lü; Qing Cui; Longfei Li; Jun Zhou; Sheng Li
• Keywords: Benchmark (surveying)

Large language models (LLMs) show promise for natural tasks but struggle when applied directly to complex domains like finance. LLMs have difficulty reasoning about and integrating all relevant information. We propose a data-centric approach enable better handle financial tasks. Our key insight is that rather than overloading the LLM with everything at once, it more effective preprocess pre-understand data. create (FLLM) using multitask prompt-based finetuning achieve data pre-processing pre-understanding. However, labeled scarce each task. To overcome manual annotation costs, we employ abductive augmentation (AAR) automatically generate training by modifying pseudo labels from FLLM's own outputs. Experiments our FLLM AAR substantially outperforms baseline designed raw text, achieving state-of-the-art on analysis interpretation also open source new benchmark interpretation. methodology provides promising path unlock LLMs' potential real-world domains.