Abstract
Tables are prevalent in real-world databases, requiring significant time andeffort for humans to analyze and manipulate. The advancements in large languagemodels (LLMs) have made it possible to interact with tables using naturallanguage input, bringing this capability closer to reality. In this paper, wepresent TableGPT, a unified fine-tuned framework that enables LLMs tounderstand and operate on tables using external functional commands. Itintroduces the capability to seamlessly interact with tables, enabling a widerange of functionalities such as question answering, data manipulation (e.g.,insert, delete, query, and modify operations), data visualization, analysisreport generation, and automated prediction. TableGPT aims to provideconvenience and accessibility to users by empowering them to effortlesslyleverage tabular data. At the core of TableGPT lies the novel concept of globaltabular representations, which empowers LLMs to gain a comprehensiveunderstanding of the entire table beyond meta-information. By jointly trainingLLMs on both table and text modalities, TableGPT achieves a deep understandingof tabular data and the ability to perform complex operations on tables throughchain-of-command instructions. Importantly, TableGPT offers the advantage ofbeing a self-contained system rather than relying on external API interfaces.Moreover, it supports efficient data process flow, query rejection (whenappropriate) and private deployment, enabling faster domain data fine-tuningand ensuring data privacy, which enhances the framework's adaptability tospecific use cases.