Hugging Face Deploys Presidio to Scan Datasets for Leaked PII

Hugging Face is rolling out an experimental feature that automatically scans datasets on the Hub for personally identifiable information, using Microsoft’s open-source Presidio tool. The move targets a persistent headache for ML practitioners: massive pre-training datasets scraped from the web that, despite filters, still contain emails, names, and other sensitive details.

The company identified two distinct categories of PII-laden datasets. The first includes annotated collections like Ai4Privacy’s PII-Masking-300k, which are intentionally built to train detection models for content moderation and anonymization. The second — and more concerning — category covers the terabyte-scale web crawl datasets used for pre-training. Hugging Face acknowledges that even with existing filtering, “small amounts of sensitive information can still slip through the cracks due to the sheer volume of data.”

Practitioners browsing the Hub will now see a report estimating PII presence in a given dataset. A sample report on one pre-training dataset flagged small quantities of emails and sensitive PII. The practical upside is straightforward: if a report surfaces red flags, a team can run additional filtering with Presidio before committing compute resources to training. Dataset owners get a different benefit — a validation layer to check their own PII scrubbing before publishing.

The stakes go beyond privacy compliance. Training on data that contains PII can cause models to learn spurious correlations between identities and outcomes, producing biased predictions or regurgitating personal information during inference. That’s a failure mode that has plagued large language models, and one that’s notoriously difficult to audit after the fact.

Hugging Face credits France’s CNIL data protection authority for guidance on GDPR compliance, noting the regulator’s updated AI how-to sheets. The Presidio reports are explicitly an experiment for now, but they signal a broader shift: dataset transparency is becoming as important as model transparency. Expect more automated auditing tools to follow as the community grapples with training data provenance.