[Re] TIES-Merging: Resolving Interference When Merging Models
Abstract: This paper presents a detailed reproduction study of the TIES-MERGING model merging technique, as introduced by Yadav et al. (2023). Our objective is to replicate the primary findings of the original research, highlighting the efficacy of TIES-MERGING over baseline models across various scenarios, including different modalities and an increased number of tasks. Through our efforts, we aim to validate these findings, assess the optimal task quantity for peak single-task model performance, and evaluate the effectiveness of employing a top-k selection process while trimming. Utilizing the codebase provided by the original authors with necessary modifications, we incorporate additional scripts for data preparation and integrate code from Yu et al. (2023) for comparison against other merging algorithms. Despite encountering some challenges, such as missing components in the provided GitHub code and a lack of responsiveness from the original authors, our results largely corroborate the claims made by Yadav et al., showing a slight deviation in performance metrics. By conducting experiments under restricted settings, this paper demonstrates that TIES-MERGING operates effectively according to our reproduction efforts, affirming its potential in model merging applications.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Simon_Kornblith1
Submission Number: 2259
Loading