Data Cleaning for Research Data Management: A Human-Supervised, LLM-Assisted Approach

Data cleaning is one of the most labour-intensive steps in research, and large language models (LLMs) are increasingly used for it, often without checking what they change. This presentation and hands-on session look at when LLMs help with data cleaning, when they silently do harm, and how to tell the difference (through provenance: tracing every change back to the step that made it). We built an agentic pipeline in which an LLM explores a dataset with code, submits cleaning code one step per problem, and gets back a single score. Every change it makes is logged and evaluated in three layers: data (what happened to the data), behaviour (did the model fix real problems, and fix them right) and provenance (can every change be traced, recomputed and undone). We ran it on the Portal rodent dataset with eight systematically injected errors, comparing models and providers. This record contains: Presentation.pdf: the slides, presented at the Research Data Conference (RDC) 2026, Bern, 20 November 2026. Hands-on.zip: the materials to try it yourself on a sample of 400 artworks from the Tate collection, with eight injected errors and an answer key.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24359396
Primary Topic
Data Quality and Management
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Data Cleaning for Research Data Management: A Human-Supervised, LLM-Assisted Approach

Sébastien Martin, Zahra Hashemi
Zenodo (CERN European Organization for Nuclear Research)
Data Quality and Management
article

Data Cleaning for Research Data Management: A Human-Supervised, LLM-Assisted Approach

Sébastien Martin, Zahra Hashemi
article en

Abstract

Data cleaning is one of the most labour-intensive steps in research, and large language models (LLMs) are increasingly used for it, often without checking what they change. This presentation and hands-on session look at when LLMs help with data cleaning, when they silently do harm, and how to tell the difference (through provenance: tracing every change back to the step that made it). We built an agentic pipeline in which an LLM explores a dataset with code, submits cleaning code one step per problem, and gets back a single score. Every change it makes is logged and evaluated in three layers: data (what happened to the data), behaviour (did the model fix real problems, and fix them right) and provenance (can every change be traced, recomputed and undone). We ran it on the Portal rodent dataset with eight systematically injected errors, comparing models and providers. This record contains: Presentation.pdf: the slides, presented at the Research Data Conference (RDC) 2026, Bern, 20 November 2026. Hands-on.zip: the materials to try it yourself on a sample of 400 artworks from the Tate collection, with eight injected errors and an answer key.

Zenodo (CERN European Organization for Nuclear Research)
Data Quality and Management
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon