While large-scale training data is fundamental for developing capable large language models (LLMs), strategically selecting high-quality data has emerged as a critical approach to enhance training efficiency and reduce computational costs. Current data selection methodologies predominantly rely on static, training-agnostic criteria, failing to account for the dynamic model training and data interactions. In this paper, we propose a new Data Weighting Model (DWM) to adjust the weight of selected data within each batch to achieve a dynamic data utilization during LLM training. Specially, to better capture the dynamic data preference of the trained model, a bi-level optimization framework is implemented to update the weighting model. Our experiments demonstrate that DWM enhances the performance of models trained with randomly-selected data, and the learned weighting model can be transferred to enhance other data selection methods and models of different sizes. Moreover, we further analyze how a model’s data preferences evolve throughout training, providing new insights into the data preference of the model during training.

Training powerful language models requires vast amounts of data, but using all available data can be inefficient, costly, and environmentally taxing — especially when much of it is low quality. While recent methods try to select better data before training begins, they often ignore how a model’s preferences change as it learns.Our work introduces a new model called the Data Weighting Model (DWM), which dynamically adjusts how much influence each piece of data has during training. Instead of treating all data in a batch equally, DWM learns to emphasize the most helpful examples at each training stage. This system is trained using a novel two-step optimization approach that tracks the model’s evolving preferences. We show that even when starting with randomly chosen data, DWM can improve training efficiency and performance — even outperforming some hand-picked datasets. The trained weighting model also works well across different model sizes and can boost other data selection techniques. Our research offers a more adaptive, cost-effective way to train large language models.