PAN-KK: A Language Resource for Plagiarism Detection in Low-Resource Kazakh

We present PAN-KK, to our knowledge, the first dataset created specifically for extrinsic plagiarism detection in Kazakh. PAN-KK is derived by translating over 20,000 English PAN 2010 plagiarism pairs and includes a human-validated subset in which 2,000 samples were reviewed by 10 native Kazakh speakers. Using PAN-KK, we train and evaluate several transformer-based models; XLM-RoBERTa Large achieves the best performance. To demonstrate practical use, we integrate the model into a two-stage detection pipeline: (i) candidate retrieval and (ii) neural verification with span-level localisation. On a manually annotated set of 1,000 Kazakh text pairs, token-based evaluation yields P=0.929, R=0.843, and F1=0.868, while span-level localisation attains P=0.718, R=0.649, and F1=0.667. PAN-KK and the accompanying pipeline provide an important starting point for improving plagiarism detection in low-resource languages such as Kazakh.