During the pretraining phase, large language models (LLMs) acquire vast amounts of knowledge from extensive text corpora. Nevertheless, in later stages such as fine-tuning and inference, the model may encounter knowledge not covered in the initial training, which can lead to hallucinations and degraded performance. This issue has a profound impact on the model's capabilities, as it will inevitably face out-of-scope knowledge after pretraining. Furthermore, fine-tuning is often required to adapt LLMs to domain-specific tasks. However, this phenomenon limits the model's ability to learn and integrate new information during fine-tuning. The effectiveness of fine-tuning largely depends on the type of knowledge involved. Existing research suggests that fine-tuning the model on partially mastered knowledge-for instance, question-answer pairs where the model has a chance of providing correct responses under non-greedy decoding-can enable the model to acquire new knowledge while mitigating hallucination. Notably, this approach can still lead to the forgetting of fully mastered knowledge, constraining the fine-tuning dataset to a narrower range and limiting the model's overall potential for improvement. Given the model's intrinsic reasoning abilities and the interconnectedness of different knowledge areas, it is likely that as the model's capacity to utilize existing knowledge improves during fine-tuning, previously unmastered knowledge may become more understandable. To explore this hypothesis, we conducted experiments and, based on the results, proposed a two-stage fine-tuning strategy. This approach not only improves the model's overall test accuracy and knowledge retention but also preserves its accuracy on previously mastered content. When fine-tuning on the WikiQA dataset, our method increases the amount of knowledge acquired by the model in this stage by 24%.

Classification on long-tailed distributed data is a challenging problem, which suffers from serious class-imbalance and hence poor performance on tail classes with only a few samples. Owing to this paucity of samples, learning on the tail classes is especially challenging for the fine-tuning when transferring a pretrained model to a downstream task. In this work, we present a simple modification of standard fine-tuning to cope with these challenges. Specifically, we propose a two-stage fine-tuning: we first fine-tune the final layer of the pretrained model with class-balanced reweighting loss, and then we perform the standard fine-tuning. Our modification has several benefits: (1) it leverages pretrained representations by only fine-tuning a small portion of the model parameters while keeping the rest untouched; (2) it allows the model to learn an initial representation of the specific task; and importantly (3) it protects the learning of tail classes from being at a disadvantage during the model updating. We conduct extensive experiments on synthetic datasets of both two-class and multi-class tasks of text classification as well as a real-world application to ADME (i.e., absorption, distribution, metabolism, and excretion) semantic labeling. The experimental results show that the proposed two-stage fine-tuning outperforms both fine-tuning with conventional loss and fine-tuning with a reweighting loss on the above datasets.