Merge to Mix: Mixing Datasets via Model Merging

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Tao, Zhixu Silvia, Vinken, Kasper, Yeh, Hao-Wei, Cooper, Avi, Boix, Xavier
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915297788690432
author Tao, Zhixu Silvia
Vinken, Kasper
Yeh, Hao-Wei
Cooper, Avi
Boix, Xavier
author_facet Tao, Zhixu Silvia
Vinken, Kasper
Yeh, Hao-Wei
Cooper, Avi
Boix, Xavier
contents Mixing datasets for fine-tuning large models (LMs) has become critical for maximizing performance on downstream tasks. However, composing effective dataset mixtures typically relies on heuristics and trial-and-error, often requiring multiple fine-tuning runs to achieve the desired outcome. We propose a novel method, $\textit{Merge to Mix}$, that accelerates composing dataset mixtures through model merging. Model merging is a recent technique that combines the abilities of multiple individually fine-tuned LMs into a single LM by using a few simple arithmetic operations. Our key insight is that merging models individually fine-tuned on each dataset in a mixture can effectively serve as a surrogate for a model fine-tuned on the entire mixture. Merge to Mix leverages this insight to accelerate selecting dataset mixtures without requiring full fine-tuning on each candidate mixture. Our experiments demonstrate that Merge to Mix surpasses state-of-the-art methods in dataset selection for fine-tuning LMs.
format Preprint
id arxiv_https___arxiv_org_abs_2505_16066
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Merge to Mix: Mixing Datasets via Model Merging
Tao, Zhixu Silvia
Vinken, Kasper
Yeh, Hao-Wei
Cooper, Avi
Boix, Xavier
Machine Learning
Artificial Intelligence
Computation and Language
Mixing datasets for fine-tuning large models (LMs) has become critical for maximizing performance on downstream tasks. However, composing effective dataset mixtures typically relies on heuristics and trial-and-error, often requiring multiple fine-tuning runs to achieve the desired outcome. We propose a novel method, $\textit{Merge to Mix}$, that accelerates composing dataset mixtures through model merging. Model merging is a recent technique that combines the abilities of multiple individually fine-tuned LMs into a single LM by using a few simple arithmetic operations. Our key insight is that merging models individually fine-tuned on each dataset in a mixture can effectively serve as a surrogate for a model fine-tuned on the entire mixture. Merge to Mix leverages this insight to accelerate selecting dataset mixtures without requiring full fine-tuning on each candidate mixture. Our experiments demonstrate that Merge to Mix surpasses state-of-the-art methods in dataset selection for fine-tuning LMs.
title Merge to Mix: Mixing Datasets via Model Merging
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2505.16066