A classic issue in the cognitive psychology of human category learning has involved the contrast between exemplar and prototype models. However, experimental tests to distinguish the models have relied almost solely on use of artificially-constructed categories composed of simplified stimuli. Here we contrast the predictions from the models in a real-world natural-science category domain--geologic rock types. Previous work in this domain used a set of complementary methods, including multidimensional scaling and direct dimension ratings, to derive a high-dimensional feature space in which the rock stimuli are embedded. The present work compares the category-learning predictions of exemplar and prototype models that make reference to this derived feature space. The experiments include conditions that should be favorable to prototype abstraction, including use of multiple large-size categories, delayed transfer testing, and real-world category structures. Nevertheless, the results of the qualitative and quantitative model comparisons point toward the exemplar model as providing a far better account of the observed results. Evidence is also provided that participants do not rely on all-or-none rote memories for the stored exemplars but rather use remembered exemplars as a basis for generalizing to novel transfer items from the learned categories. Limitations and directions of future work are discussed.