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

Hyunwoo Kim, Niloofar Mireshghallah, Michael Duan, Rui Xin, Shuyue Stella Li, Jaehun Jung, David Acuna, Qi Pang, Hanshen Xiao, G. Edward Suh, Sewoong Oh, Yulia Tsvetkov, Pang Wei Koh, Yejin Choi
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:58549-58593, 2026.

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. 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.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-kim26az, title = {Privasis: Synthesizing the Largest "{P}ublic" Private Dataset from Scratch}, author = {Kim, Hyunwoo and Mireshghallah, Niloofar and Duan, Michael and Xin, Rui and Li, Shuyue Stella and Jung, Jaehun and Acuna, David and Pang, Qi and Xiao, Hanshen and Suh, G. Edward and Oh, Sewoong and Tsvetkov, Yulia and Koh, Pang Wei and Choi, Yejin}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {58549--58593}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/kim26az/kim26az.pdf}, url = {https://proceedings.mlr.press/v306/kim26az.html}, 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. 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.} }
Endnote
%0 Conference Paper %T Privasis: Synthesizing the Largest "Public" Private Dataset from Scratch %A Hyunwoo Kim %A Niloofar Mireshghallah %A Michael Duan %A Rui Xin %A Shuyue Stella Li %A Jaehun Jung %A David Acuna %A Qi Pang %A Hanshen Xiao %A G. Edward Suh %A Sewoong Oh %A Yulia Tsvetkov %A Pang Wei Koh %A Yejin Choi %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-kim26az %I PMLR %P 58549--58593 %U https://proceedings.mlr.press/v306/kim26az.html %V 306 %X 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.
APA
Kim, H., Mireshghallah, N., Duan, M., Xin, R., Li, S.S., Jung, J., Acuna, D., Pang, Q., Xiao, H., Suh, G.E., Oh, S., Tsvetkov, Y., Koh, P.W. & Choi, Y.. (2026). Privasis: Synthesizing the Largest "Public" Private Dataset from Scratch. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:58549-58593 Available from https://proceedings.mlr.press/v306/kim26az.html.

Related Material