# AutoDCWorkflow: LLM-based Data Cleaning Workflow Auto-Generation and Benchmark

**Authors:** Lan Li, Liri Fang, Vetle I. Torvik  
**arXiv:** [2412.06724](https://arxiv.org/abs/2412.06724)  
**Date:** 2024-12-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2412.06724v1/

## Key points

- Large Language Models (LLMs) used in automating data-cleaning workflows through AutoDCWorkflow
- Three common data quality issues addressed: duplicates, missing values, inconsistent data formats
- Selecting target columns related to the purpose
- Inspecting data quality in each target column and generating a Data Quality Report
- Predicting the next operation and its arguments based on data quality assessment results
- Introduction of a data cleaning benchmark with annotated datasets, raw tables, clean tables, workflows, and answer sets for evaluating LLM agents' ability
- AutoDCWorkflow aims to automate real-world data cleaning operations using LLMs effectively and record these workflows for future reference
- Exploration of LLMs as automated data cleaning agents capable of generating tailored workflows fit for specific purposes
- Future research focus on examining column schema dependencies and interdependencies among different data cleaning operations to enhance precision and effectiveness

## Summary

We investigate the capabilities of Large Language Models (LLMs) in automating data-cleaning workflows through our LLM-based Auto Data Cleaning Workflow (AutoDCWorkflow). This pipeline utilizes LLMs to address three common data quality issues: duplicates, missing values, and inconsistent data formats. Our goal is to generate a clean table from a dirty one that fulfills a specific purpose expressed as a query. The process involves three key components driven by LLMs: selecting target columns related to the purpose, inspecting the quality of data in each target column and generating a Data Quality Report, and predicting the next operation and its arguments based on the results of the data quality assessment. To evaluate LLM agents' ability to automatically generate workflows for different levels of difficulty, we introduce a data cleaning benchmark with annotated datasets, raw tables, clean tables, data cleaning workflows, and answer sets. Building upon previous work, our AutoDCWorkflow aims to analyze how effectively LLMs can automate real-world data cleaning operations and record these workflows for future reference. We explore the potential of LLMs as automated data cleaning agents capable of generating tailored workflows fit for specific purposes. This study represents an initial attempt at fully automating the process of generating data cleaning workflows using LLMs. In future research, we will focus on examining column schema dependencies and interdependencies among different data cleaning operations in automated workflow generation to enhance precision and effectiveness. Overall, our findings suggest that LLMs show promise in inferring necessary cleaning steps and applying them accurately across diverse datasets without requiring fine-tuning.