Procedural content generation via machine learning (PCGML) has been demonstrating its usefulness as a content and game creation approach, and has been shown to be able to support human creativity. In this paper we present an example-driven adaptation of a classic PCG approach, binary space partition (BSP), that takes a structural template or sketch of a level and fills in the details from examples. We show that this example-driven adaptation can generate a diverse set of levels from a single structural template. We evaluate the levels generated in terms of difference between paths through the levels, amount of the level copied from the examples, and other common PCG level evaluation metrics. Furthermore, we compare this method to a Markov chain approach and show that our BSP approach matches the training level distribution better while generating a greater range of interesting features.