Reliable datasets are essential for developing and evaluating the models for fake news detection. Although the research on disinformation detection has advanced significantly in high-resource languages such as English and Chinese, Polish remains an under-resourced language in this area despite the emergence of first dedicated datasets. This paper introduces the POLfake – new relational dataset for Polish-language fake news detection. The dataset comprises nearly 6,000 fact-checked claims and 1,500 tweets that explicitly propagate these claims. The relational structure of the dataset reflects real-world misinformation dynamics by linking posts to the specific claims they disseminate. Each element is annotated with one of five levels of veracity, allowing for both fine-grained and binary classification setups. The POLfake dataset also includes a rich set of attributes for both claims and tweets such as publication time, author information and popularity metrics, enabling a wide range of approaches to development of fake news detection models. To support reproducibility and fair evaluation, two benchmark tasks are provided: Claim-Tweet Challenge and Claim-Only Challenge. POLfake dataset is intended to serve as a valuable resource for advancements in the field of detecting the Polish-language fake news.