How to check a dataset for sensitive data before training
Models remember what they are trained on. A customer email, a date of birth or a salary figure that slips into a training set can surface later in a model's output, in a log, or in a dataset someone else downloads. The check takes a minute and belongs before any of that happens — before fine-tuning, before sharing a file with a colleague, and before publishing a dataset on a hub.
To use it:
- Drop a .csv file (up to 5 MB) or paste the CSV text with its header row. Comma, semicolon, tab or pipe separated — the delimiter is detected from the first line.
- Press Check dataset. The checker samples up to 1,000 rows evenly across the file, then matches every column two ways: by its name (email, date_of_birth, salary, diagnosis) and by the shape of its values (email addresses, phone numbers, card numbers checked with the Luhn algorithm, IBANs checked with mod-97, IP addresses, dates, postcodes).
- Work through the Sensitive columns first. Each finding shows what was detected, how much of the column matches, a masked example, and a recommended treatment — remove, hash, generalise or pseudonymise.
- Check the quasi-identifier warning. Postcode plus age plus gender can single out a person even after the direct identifiers are gone. Generalise or drop at least one of each pair.
- Download the report (Markdown or JSON) and keep it with the dataset as your audit trail.
What makes this checker different
The free tools that rank for this search are scrubbers, not checkers: PrivHaven scrubs CSVs in the browser but caps its free tier at 10,000 rows, keeps its reproducible report behind Pro, and does not detect names or addresses at all; the Gera Tools Data Anonymizer replaces values with fakes but produces no audit report and no per-column risk verdict; the Unfckr CSV PII Scrubber masks, hashes or drops columns but likewise covers only emails, phones, SSNs and cards, with no names, health or financial columns and no report. The developer options — Microsoft Presidio, privacy-scan-ml and releaseguard — are libraries and command-line tools that need installing and a Python environment, and PII Crawler is a $497 desktop app. This page is the free, no-account, no-upload checker built for the AI-training question specifically: per-column verdicts across 24 sensitive types, quasi-identifier warnings, masked samples, and a downloadable audit report you can keep with the dataset.
When is this the right check?
Whenever a dataset is about to leave your machine with people in it: before fine-tuning or training a model, before sharing a file with a colleague or a contractor, before publishing a dataset, and before pasting rows into an AI chat. It is also the fast first pass for GDPR and EU AI Act dataset governance — a record of what was checked and what was flagged, in the report you download.
Honest limitations
Pattern-based, not a guarantee
Detection matches column names and value shapes. A clean result means no sensitive pattern was found in the sampled rows — it does not prove the dataset is safe. Spot-check the data yourself before publishing or training on it.
Names come from headers only
Values alone cannot reliably identify a name, so names are detected from column names such as first_name or full_name. A column called "col_7" full of names will not be flagged — rename columns honestly before scanning.
No free-text entity recognition
A long notes column full of names, addresses and phone numbers in running text is not scanned as entities. If a column holds free text about people, treat it as sensitive and review it by eye.
Only a sample is scanned
Up to 1,000 rows, sampled evenly. A sensitive value that appears only in row 50,000 will not be seen. For a full check, scan the file in extracts.
It checks, it does not clean
This tool flags sensitive columns and tells you what to do with each one — it does not redact, mask or rewrite your file. Apply the recommendations in your own pipeline, then re-check the cleaned file.
Not legal advice
The verdicts and recommendations are a practical first pass, not a compliance opinion. GDPR, the EU AI Act and sector rules have their own requirements — check them, or ask a qualified adviser, before publishing personal data.