Evolutionary Optimisation of Model Merging Recipes

23rd August 2024Last Updated 19th September 2024

arXiv

• Introduction

  Introduces evolutionary optimisation of model merging recipes to evolve pre-trained finetuned models at low computational cost

• Motivation

  Unlock new capabilities by simple merging of abundantly available fine-tuned models without gradient-based optimisation

• Research Question

  Can classical model merging techniques be automated via evolutionary algorithms?

• Methodology

  Defines a search space of model merging recipes in parameter space (PS) and data flow space (DFS) and applies evolutionary algorithms to search for optimal recipes

• Results

  Merged general Japanese and math model surpasses parent models on Japenese math benchmarks