Offline signature verification is a challenging task for both computer science and forensics. Skilled forgeries often cannot be recognized by humans, which leads to the need to develop automated forged signatures recognition methods, which in turn requires the creation of different datasets for training models, which include the NSP – the first dataset with Cyrillic offline signatures, including genuine signatures with their skilled and simple forgeries. The process of collecting data for this dataset is described in detail. In the process of collecting samples we reformulated the forensic classification of signatures by criterion of their structure and forgery vulnerability. Gathered database was evaluated using a Siamese neural network model and the results are compared with the same model trained on CEDAR dataset.