Privasis: Synthesizing the Largest "Public" Private Dataset from Scratch

Research involving privacy-sensitive data has always been constrained by data scarcity, standing in sharp contrast to other areas that have benefited from data scaling. To quench this thirst, we present Privasis (i.e., privacy oasis), the first million-scale fully synthetic dataset entirely built from scratch—an expansive reservoir of texts with rich and diverse private information—designed to broaden and accelerate research in areas where processing sensitive social data is inevitable. Compared to existing datasets, Privasis, comprising 1.4 million records, offers orders-of-magnitude larger scale with quality, and far greater diversity across various document types, including medical records, legal documents, financial records, calendars, emails, meeting transcripts, and text-messages with a total of 55.1 million annotated attributes such as ethnicity, date of birth, workplace, etc. We leverage Privasis to construct a parallel corpus for text sanitization with our pipeline that recursively decomposes texts and applies targeted sanitization. Our compact sanitization models ($\leq$ 4B) trained on this dataset outperform state-of-the-art large language models, such as GPT-5 and Qwen-3 235B.

AI assistants now handle our personal documents—medical records, emails, bank statements—but research on protecting privacy in these systems has lagged for years. Real private data can't legally be shared, leaving researchers stuck with tiny datasets while AI elsewhere surges ahead on massive ones. To break this bottleneck, we built Privasis ("privacy oasis"), the first million-scale collection of realistic private documents that contains no real people's information. Its 1.4 million records—spanning medical visits, legal filings, bank statements, calendars, and text messages—were generated by AI entirely from scratch, without using any real document. We then trained small, lightweight models that read a document and selectively remove or blur sensitive details on request, for example turning "March 3rd" into "early spring." Despite being a fraction of their size, our models outperform today's largest systems like GPT-5 at this task. Privasis removes a long-standing roadblock for privacy research. Because our models are small enough to run on a personal device, sensitive text never needs to leave the user's machine, which is an important step in the age of AI assistants. We release all data, models, and code openly so the community can build AI that respects privacy by design.