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

TL;DR

Privasis tackles the data-scarcity barrier in privacy research by introducing a million-scale, fully synthetic dataset of privacy-rich records generated from scratch. It couples this with Privasis-Sanitization, a decomposition-based, instruction-driven sanitization pipeline and a suite of lightweight sanitizers trained on the synthetic corpus, enabling on-device privacy-preserving processing. The work demonstrates high lexical and semantic diversity across domains, and shows that compact models (≤4B) can outperform frontier LLMs on vanilla sanitization tasks while maintaining robust retention of non-sensitive content. It also provides a rigorous leakage- and retention-focused evaluation framework, along with strong privacy safeguards and a commitment to open science, to accelerate future research in privacy-aware AI agents and sanitization methods.

Abstract

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. This challenge is becoming increasingly urgent as modern AI agents--such as OpenClaw and Gemini Agent--are granted persistent access to highly sensitive personal information. To tackle this longstanding bottleneck and the rising risks, 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 history, legal documents, financial records, calendars, 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 decomposes texts and applies targeted sanitization. Our compact sanitization models (<=4B) trained on this dataset outperform state-of-the-art large language models, such as GPT-5 and Qwen-3 235B. We plan to release data, models, and code to accelerate future research on privacy-sensitive domains and agents.

Figures (5)

• Figure 1: Privasis, the Privacy Oasis Dataset: We synthesize the first publicly-available million-scale dataset with diverse private information, entirely from scratch. (a) Using auxiliary control variables, we initialize a text record draft containing rich private information and then selectively refine it while preserving the overall diversity of the record set (§ \ref{['sec:synthesis']}). (b) Based on this, we construct a parallel corpus for text sanitization using a decomposition-based sanitization pipeline (§ \ref{['sec:sanitization']}). (c) On this parallel corpus, we train compact sanitization models ($\leq$4B) that outperform GPT-5 (§ \ref{['sec:experiments']}).
• Figure 2: Overview of our synthesis pipeline.
• Figure 3: The ethnicity distribution in Privasis.
• Figure 4: Distribution of the subcategories of the records in Privasis-Sanitization.
• Figure 5: Sanitization performance for different chunk sizes.