Automated software engineering is increasingly constrained by the quality and structure of training data rather than by model capacity. This challenge is particularly pronounced in automated code summarization, with Class-Level Code Summarization being a prime example: object-oriented dependencies introduce substantial noise into mined repositories, and the resulting documentation quality varies widely. Moreover, practical compute budgets often restrict training to only a fraction of available data, making dataset selection a first-order design decision. However, existing approaches typically treat dataset construction as a pre-processing step, offering limited guidance on how to systematically balance data quality and coverage under fixed training budgets. To address this gap, we propose EvoQD (Evolutionary Quality–Diversity Selection), a search-based framework for selecting fixed-budget training subsets for summarizing higher-order code components, particularly classes, from large code corpora. EvoQD casts dataset selection as a quality–diversity optimization problem and combines: (1) an LLM-as-a-Judge that rates adequacy, internal consistency, and architectural insight; (2) a multi-attribute quality model enriched with code-centric signals such as validity, complexity, alignment, and redundancy; and (3) a genetic algorithm that searches for fixed-budget subsets maximizing both aggregate quality and semantic coverage.