Research
data-cleaner is the state of the art in real-time data-cleaning technologies.
data-cleaner is built and maintained by Edward Gaere at ETH Zurich. The research behind it, with Prof. Dr. Florian von Wangenheim, creates new benchmarks and real-time cleaning models for real-world data as it actually appears, not after it has been tidied. The benchmarks measure how leading machine-learning and language models handle that data. data-cleaner is designed to handle the majority of surface forms arising in practice at the start of a data stack: bytes, encodings, file formats, shapes, locales, aliases, permutations, and so on.
The MESSY series
Benchmarks on real-world data: addresses, datetimes and files.
Addresses: MESSY STREETS
MESSY STREETS: A Benchmark for Geocoding Real-World Addresses. Edward Gaere and Florian von Wangenheim. Accepted as a short paper at ACM SIGSPATIAL 2026, Riverside, 3 to 6 November 2026.
A benchmark of addresses taken verbatim from the web, with parts that are missing, repeated or malformed, and with reference locations from OpenAddresses and OpenStreetMap. Twelve commercial and open-source geocoders are compared on it. Commercial geocoders find far more of these addresses, and most of the gap comes from addresses that are not written in canonical form. That is the case for cleaning an address before geocoding it.
- Preprint on arXiv: 2609.01612
- DOI: 10.48550/arXiv.2609.01612
- ACM SIGSPATIAL 2026: accepted short papers
Datetimes: MESSY TIMES
PRIMETIME: Limits of LLMs in Temporal Primitives. Edward Gaere and Florian von Wangenheim. The preprint of MESSY TIMES.
Language models are tested on each basic datetime operation in isolation. The operations are individually unreliable across models, and learnable by fine-tuning on generated data.
Files: MESSY CSV
Forthcoming.
Clean a file · MCP · API · Research · llms.txt · REST reference · OpenAPI · Data policy