Saved in:
Bibliographic Details
Main Authors: McKenna, Ryan, Huang, Yangsibo, Sinha, Amer, Balle, Borja, Charles, Zachary, Choquette-Choo, Christopher A., Ghazi, Badih, Kaissis, George, Kumar, Ravi, Liu, Ruibo, Yu, Da, Zhang, Chiyuan
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2501.18914
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909471067865088
author McKenna, Ryan
Huang, Yangsibo
Sinha, Amer
Balle, Borja
Charles, Zachary
Choquette-Choo, Christopher A.
Ghazi, Badih
Kaissis, George
Kumar, Ravi
Liu, Ruibo
Yu, Da
Zhang, Chiyuan
author_facet McKenna, Ryan
Huang, Yangsibo
Sinha, Amer
Balle, Borja
Charles, Zachary
Choquette-Choo, Christopher A.
Ghazi, Badih
Kaissis, George
Kumar, Ravi
Liu, Ruibo
Yu, Da
Zhang, Chiyuan
contents Scaling laws have emerged as important components of large language model (LLM) training as they can predict performance gains through scale, and provide guidance on important hyper-parameter choices that would otherwise be expensive. LLMs also rely on large, high-quality training datasets, like those sourced from (sometimes sensitive) user data. Training models on this sensitive user data requires careful privacy protections like differential privacy (DP). However, the dynamics of DP training are significantly different, and consequently their scaling laws are not yet fully understood. In this work, we establish scaling laws that accurately model the intricacies of DP LLM training, providing a complete picture of the compute-privacy-utility tradeoffs and the optimal training configurations in many settings.
format Preprint
id arxiv_https___arxiv_org_abs_2501_18914
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Scaling Laws for Differentially Private Language Models
McKenna, Ryan
Huang, Yangsibo
Sinha, Amer
Balle, Borja
Charles, Zachary
Choquette-Choo, Christopher A.
Ghazi, Badih
Kaissis, George
Kumar, Ravi
Liu, Ruibo
Yu, Da
Zhang, Chiyuan
Machine Learning
Cryptography and Security
Scaling laws have emerged as important components of large language model (LLM) training as they can predict performance gains through scale, and provide guidance on important hyper-parameter choices that would otherwise be expensive. LLMs also rely on large, high-quality training datasets, like those sourced from (sometimes sensitive) user data. Training models on this sensitive user data requires careful privacy protections like differential privacy (DP). However, the dynamics of DP training are significantly different, and consequently their scaling laws are not yet fully understood. In this work, we establish scaling laws that accurately model the intricacies of DP LLM training, providing a complete picture of the compute-privacy-utility tradeoffs and the optimal training configurations in many settings.
title Scaling Laws for Differentially Private Language Models
topic Machine Learning
Cryptography and Security
url https://arxiv.org/abs/2501.18914