Model Merging in Pre-training of Large Language Models
This work addresses the problem of inefficient and costly pre-training for large language model developers, though it appears incremental as it builds on existing model merging techniques.
The paper investigates model merging techniques during large-scale pre-training of language models, demonstrating that merging checkpoints trained with constant learning rates achieves significant performance improvements and enables accurate prediction of annealing behavior across models from millions to over 100 billion parameters.
Model merging has emerged as a promising technique for enhancing large language models, though its application in large-scale pre-training remains relatively unexplored. In this paper, we present a comprehensive investigation of model merging techniques during the pre-training process. Through extensive experiments with both dense and Mixture-of-Experts (MoE) architectures ranging from millions to over 100 billion parameters, we demonstrate that merging checkpoints trained with constant learning rates not only achieves significant performance improvements but also enables accurate prediction of annealing behavior. These improvements lead to both more efficient model development and significantly lower training costs. Our detailed ablation studies on merging strategies and hyperparameters provide new insights into the underlying mechanisms while uncovering novel applications. Through comprehensive experimental analysis, we offer the open-source community practical pre-training guidelines for effective model merging.