Supervised fine-tuning (SFT) is crucial for adapting Large Language Models (LLMs) to specific tasks. In this work, we demonstrate that the order of training data can lead to significant training imbalances, potentially resulting in performance degradation. Consequently, we propose to mitigate this imbalance by merging SFT models fine-tuned with different data orders, thereby enhancing the overall effectiveness of SFT. Additionally, we introduce a novel technique, “parameter-selection merging,” which outperforms traditional weighted-average methods on five datasets. Further, through analysis and ablation studies, we validate the effectiveness of our method and identify the sources of performance improvements.

Commonsense question answering is important for making decisions about everyday matters. Although existing commonsense question answering works based on fully fine-tuned PLMs have achieved promising results, they suffer from prohibitive computation costs as well as poor interpretability. Some works improve the PLMs by incorporating knowledge to provide certain evidence, via elaborately designed GNN modules which require expertise. In this paper, we propose a simple knowledgeable parameter efficient tuning network to couple PLMs with external knowledge for commonsense question answering. Specifically, we design a trainable parameter-sharing adapter attached to a parameter-freezing PLM to incorporate knowledge at a small cost. The adapter is equipped with both entity- and query-related knowledge via two auxiliary knowledge-related tasks (i.e., span masking and relation discrimination). To make the adapter focus on the relevant knowledge, we design gating and attention mechanisms to respectively filter and fuse the query information from the PLM. Extensive experiments on two benchmark datasets show that KPE is parameter-efficient and can effectively incorporate knowledge for improving commonsense question answering.