Complex event processing (CEP) matches patterns over a continuous stream of events to detect situations of interest. Yet, the definition of an event pattern that precisely characterises a particular situation is challenging: there are manifold dimensions to correlate events, including time windows and value predicates. In the presence of historic event data that is labelled with the situation to detect, event patterns can be learned automatically. To cope with the combinatorial explosion of pattern candidates, existing approaches work on a type-level and discover patterns based on predefined event abstractions, aka event types. Hence, discovery is limited to patterns of a fixed granularity and users face the burden to manually select appropriate event abstractions. We present IL-M iner , a system that discovers event patterns by genuinely working on the instance-level, not assuming a priori knowledge on event abstractions. In a multi-phase process, IL-M iner first identifies relevant abstractions for the construction of event patterns. The set of events explored for pattern discovery is thereby reduced, while still providing formal guarantees on correctness, minimality, and completeness of the discovery result. Experiments using real-world datasets from diverse domains show that IL-M iner discovers a much broader range of event patterns compared to the state-of-the-art in the field.