<p>Obtaining sufficient data is often challenging, costly, or even infeasible in real-world applications. To address this issue, researchers have developed methods for learning from limited data, which involves discovering the underlying pattern using a small amount of available data. Data augmentation and model fine-tuning are two methods commonly used for learning with limited data. However, both methods have limitations: model fine-tuning methods often face difficulties in acquiring relevant data for pre-training in certain fields, such as drug discovery, whereas data augmentation methods have the potential risk of causing bias. To address this challenge, we propose a general learning framework for limited data using a combination of data augmentation and model fine-tuning (DAMFT). This framework comprises data augmentation and model fine-tuning modules. The data augmentation module is responsible for generating relevant data that are utilized to pre-train the target model, and the model fine-tuning module aims to correct the potential bias of the pre-trained model caused by data augmentation. DAMFT not only solves the difficulty in obtaining relevant data needed for model fine-tuning but also alleviates the bias caused by data augmentation. Furthermore, under the DAMFT framework, we provide a supervised learning algorithm, DAMFT_GH, which adopts a generative adversarial network and head fine-tuning to generate the new instances and tune the pre-trained model, respectively. The experimental results demonstrate that the proposed learning algorithm can effectively improve the performance of the classification model in the case of limited data.</p>