Cognitive neuroscientists are trained to use clever experimental manipulations to decompose complex cognitive phenomena—like language—into simpler, more manageable components. This approach relies on the assumption that we can recompose our piecemeal models into a satisfying understanding of ecological brain function. Drawing parallels with machine learning and evolutionary theory, we challenge this assumption and argue that ecological considerations should play a more central role in cognitive neuroscience. One critical step in this direction is the development of well-curated public datasets that can serve as benchmarks for model evaluation. With this in mind, we present the “Narratives” data collection—a naturalistic fMRI dataset comprising over 800 scans acquired while participants listened to a variety of spoken stories. We discuss both opportunities and challenges that arise in scaling up to a more community-based paradigm for model development in naturalistic neuroimaging.