Music Information Retrieval techniques provide a useful approach for identifying folk music, supporting the preservation of intangible cultural heritage by accurately recognising characteristic songs. Artificial intelligence-enabled music detection remains an intricate issue. Thus, we propose a novel neural network architecture for Polish folk music classification, a CNN-based model with an Adaptive Attention Module (CNN AAM). It comprises an adaptive trunk branch, an adaptive mask branch, and an adaptive gate, allowing the model to adjust the attention mechanism based on the most relevant sound features. Furthermore, the proposed approach addresses challenges related to limited and imbalanced folk music datasets. As an input for the proposed architecture, Mel spectrograms, spectrograms, scalograms, and Mel-Frequency Cepstral Coefficients plots have been gathered, demonstrating various features of 3-second audio recordings. This study uses a dataset of music from Polish national dances, including the Polonez, Oberek, Mazur, Krakowiak, and Kujawiak. The obtained accuracy results are encouraging, reaching over 95%. The effectiveness of the developed model is compared with the state-of-the-art classifiers. Moreover, the impact of each input is verified during ablation studies. Our approach demonstrates that the CNN AAM demonstrates improved performance compared with typical content-based music models.