Table-GPT is a language model fine-tuned for diverse table tasks. The authors propose a new "table fine-tuning" approach, where language models like GPT-3.5 and ChatGPT are trained using diverse table tasks synthesized from real tables. This approach aims to enhance language models' ability to understand and perform table-related tasks. The resulting TABLE-GPT models outperform vanilla GPT-3.5 and ChatGPT on a wide range of table tasks, including data transformation, data cleaning, data imputation, and table-QA. They also demonstrate strong generalizability, responding to new and unseen table tasks in a manner similar to GPT-3.5 and ChatGPT. The authors release their code, training data, and an extensive evaluation benchmark for future research. The paper highlights the limitations of current language models in understanding tables, as they are pre-trained on one-dimensional natural-language texts, while tables are two-dimensional. The authors propose a "table-tuning" paradigm, analogous to instruction-tuning, to improve language models' ability to understand tables. They synthesize diverse table tasks from real tables and augment them at instruction, table, and completion levels to increase data diversity. The resulting TABLE-GPT models show improved performance on table tasks and generalizability to new tasks. The paper also discusses the importance of vertical reading in tables for many common tasks and the sensitivity of language models to the order of rows and columns in tables. The authors conclude that TABLE-GPT is a promising approach for improving language models' ability to understand and perform table tasks.Table-GPT is a language model fine-tuned for diverse table tasks. The authors propose a new "table fine-tuning" approach, where language models like GPT-3.5 and ChatGPT are trained using diverse table tasks synthesized from real tables. This approach aims to enhance language models' ability to understand and perform table-related tasks. The resulting TABLE-GPT models outperform vanilla GPT-3.5 and ChatGPT on a wide range of table tasks, including data transformation, data cleaning, data imputation, and table-QA. They also demonstrate strong generalizability, responding to new and unseen table tasks in a manner similar to GPT-3.5 and ChatGPT. The authors release their code, training data, and an extensive evaluation benchmark for future research. The paper highlights the limitations of current language models in understanding tables, as they are pre-trained on one-dimensional natural-language texts, while tables are two-dimensional. The authors propose a "table-tuning" paradigm, analogous to instruction-tuning, to improve language models' ability to understand tables. They synthesize diverse table tasks from real tables and augment them at instruction, table, and completion levels to increase data diversity. The resulting TABLE-GPT models show improved performance on table tasks and generalizability to new tasks. The paper also discusses the importance of vertical reading in tables for many common tasks and the sensitivity of language models to the order of rows and columns in tables. The authors conclude that TABLE-GPT is a promising approach for improving language models' ability to understand and perform table tasks.