Multi-label text classification (MLTC) presents a complex challenge, requiring classifiers to assign multiple labels to each document, resulting in diverse label combinations across a large corpus. While substantial advancements have been made in traditional MLTC tasks, existing studies often overlook the impact of compositional generalization (CG) challenges on model performance. Current models struggle to generalize to novel and complex label combinations, which are typically formed from fundamental labels. This study addresses this limitation by proposing a straightforward and scalable data grouping approach, termed best-fit document (BFD), which optimizes large-scale training documents based on their lengths to enhance the classifier’s capacity to model novel label combinations. A comprehensive evaluation of BFD was conducted across four datasets with specialized data splits. The experimental findings indicate that our method significantly enhances the CG ability of classification models on the evaluated benchmarks.