# ChatPipe: Orchestrating Data Preparation Program by Optimizing Human-ChatGPT Interactions

**Authors:** Sibei Chen, Hanbing Liu, Weiting Jin, Xiangyu Sun, Xiaoyao Feng, Ju Fan, Xiaoyong Du, Nan Tang  
**arXiv:** [2304.03540](https://arxiv.org/abs/2304.03540)  
**Date:** 2023-04-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.03540v1/

## Key points

- Authors address the need for a high-quality data preparation program in successful machine learning endeavors
- Acknowledge limitations despite capabilities of large language models like ChatGPT
- Emphasize the importance of providing specific prompts to guide ChatGPT iteratively
- Introduce ChatPipe as a system to facilitate interactions between users and ChatGPT
- ChatPipe offers recommendations on next data preparation operations and allows easy reverting to previous program versions
- Practical demonstrations of ChatPipe's capabilities through real-world ML tasks from Kaggle
- Optimizing human-ChatGPT interactions through ChatPipe can enhance ML workflows and achieve more robust outcomes

## Summary

In their paper titled "ChatPipe: Orchestrating Data Preparation Program by Optimizing Human-ChatGPT Interactions," authors Sibei Chen, Hanbing Liu, Weiting Jin, Xiangyu Sun, Xiaoyao Feng, Ju Fan, Xiaoyong Du, and Nan Tang address the crucial need for a high-quality data preparation program in successful machine learning (ML) endeavors. They acknowledge the time and effort-consuming nature of this process and highlight the limitations faced despite the remarkable capabilities of large language models like ChatGPT in generating programs through natural language prompts. The authors emphasize that users must provide specific prompts to guide ChatGPT iteratively in enhancing data preparation programs. This requirement necessitates a certain level of expertise in programming, familiarity with the dataset being used, and understanding of the ML task at hand. Furthermore, once a program is generated, revisiting a previous version or making modifications to it without restarting the entire process poses challenges. To address these issues, the authors introduce ChatPipe as a novel system designed to facilitate seamless interactions between users and ChatGPT. ChatPipe offers effective recommendations on next data preparation operations and guides ChatGPT in generating programs for these operations. Additionally, ChatPipe enables users to easily revert to previous versions of the program, streamlining experimentation and testing processes for increased efficiency. The development of a web application for ChatPipe along with the preparation of several real-world ML tasks from Kaggle serve as practical demonstrations of its capabilities. These tasks not only showcase how ChatPipe can expedite the orchestration of high-quality data preparation programs but also provide VLDB attendees with an opportunity to explore its innovative features firsthand. By optimizing human-ChatGPT interactions through ChatPipe  and  can enhance their ML workflows and achieve more robust outcomes in their projects.