CollabEdit: Towards Non-destructive Collaborative Knowledge Editing

Jiamu Zheng, Jinghuai Zhang, Futing Wang, Tianyu Du, Tao Lin

Keywords: Collaborative Learning, Knowledge Editing

Abstract:

Recently, collaborative fine-tuning large language models (LLMs) has emerged
as a new paradigm for utilizing private data from different parties in a manner
that guarantees both efficiency and privacy. Meanwhile, the practical needs of the
“right to be forgotten” and the frequent demands to update outdated information,
have led to a burgeoning in the techniques of knowledge editing (KE) for LLMs.
However, current KE methods are all designed for a single model, and directly
adapting current KE methods to collaborative learning scenarios encounters severe
performance decreases. In this study, we propose a non-destructive collaborative
knowledge editing framework COLLABEDIT that utilizes novel model fusion strategy to preserve overall editing performance. Empirical studies on two canonical
datasets demonstrate the effectiveness and superiority of our method compared
with other destructive baselines.

Submission Number: 25