Data Quality

Quality depends on the job. A customer table where 10% of email addresses are missing may be fine for a revenue report and useless for an email campaign. A support-ticket archive full of duplicates may still be good for search but will skew a model trained to predict ticket volume. So the useful question is not “is our data clean?” but “is it good enough for this decision?”

For AI, problems show up in two places. In training data, mislabeled, duplicated or unrepresentative examples teach a model the wrong lesson, and more of them rarely fixes that. In the data an application reads at run time, such as the documents behind RAG, outdated or contradictory records produce confident wrong answers that look like hallucinations. Google researchers call the compounding version a “data cascade”: a data issue early in a project that surfaces as a failure much later.

Practical fixes are mostly unglamorous: validation checks inside each data pipeline, clear owners and definitions in a data catalog, deduplication, and documentation such as datasheets that say what a dataset is and is not for.

Guests treat it as the core of model work. In episode 43 at 25:34, Liquid AI’s Maxime Labonne calls defining data quality “the most important question” in training models, adding that training techniques “will never replace good data.” In episode 2 at 4:59, Weaviate’s Bob van Luijt names low data quality as “still the number one problem” in data management. Episode 2 is a Season 1 panel sponsored by Galileo and hosted by Galileo co-founder Yash Sheth (intro at 0:45).