1 paper
Yuzheng Hu, Fan Wu, Ruicheng Xian +5
We propose the notion of empirical privacy variance and study it in the context of differentially private fine-tuning of language models. Specifically, we show that models calibrat…