A generative data model is a simplified and idealized understanding of an observation, capable of generating data values. Its purpose is to derive data through theoretical or real observation and to generate samples. These are models with structure and parameters, as opposed to holistic expert judgments based on intuition, debatable beliefs and conjectures. HRA data are often based on judged samples and collected by hand entirely, or semi-automatically at best. This obscures generality, impairs availability and potentially violates privacy. Depending on the safety analysis domain, many generative data models are “side projects” as part of thermal-hydraulic simulations, simulator-based validation or implementation of procedures during training sessions. The generative HRA data models could improve quality by obtaining a link between models' parameter input and output for their statistical and structural analyses. In this way it is also possible to prove the correctness and completeness of a HRA technique without introducing statistical biases through samples with questionable representativeness. The paper presents the capacities of the Performance Evaluation of Teamwork (PET) method and its generative data models for HRA data mining. These models give opportunity for generating data and creating of database for failure events of socio-technical system based on correct symptom definitions, symptom-based context and context-sensitive reliability evaluation of system's behaviour. They could also address the emerging issues and challenges of designing, testing, benchmarking, and evaluating HRA data and techniques.