A system for dictionary generation that can generate a set of candidate dictionaries based at least in part on subsets of content, where candidate dictionaries of the set of candidate dictionaries are generated based at least in part on a different subset of the content. The system can further use candidate dictionaries of the set of candidate dictionaries to compress the content and can identify one or more dictionary quality metrics for candidate dictionaries of the set of candidate dictionaries based at least in part on the respective compression of the content by candidate dictionaries of the set of candidate dictionaries.