Vision-language sample scoring with dynamic coverage-centric strategies

Abstract

Data pruning reduces image-classifier training cost by replacing a full training set with a coreset. Existing methods often require training a target or proxy model to estimate sample importance, which ties data selection to a particular architecture. In addition, common allocation rules do not explicitly change the selected score distribution with the retention budget. We study how to select a fixed-size subset before target-model training while enabling score reuse across architectures and budget-dependent allocation. For this image-classification data-pruning task, we use a frozen Contrastive Language–Image Pre-training (CLIP) encoder as the artificial-intelligence component of a two-stage framework. CLIPSelector computes a two-prompt image-label alignment probability for each training image without optimizing a target or proxy classifier. Dynamic Coverage-centric Coreset Selection (DCCS) partitions the ranked scores into equal-count strata, allocates the initial budget to higher-scoring strata, and samples the remaining budget from a merged lower-scoring region using importance-weighted and uniform probabilities. Experiments on three image-classification benchmarks show that CLIPSelector-DCCS achieves the best average rank among the evaluated methods, generally improves on coverage-centric allocation under matched scoring functions, supports score reuse across target architectures and retention budgets, and reduces downstream training time. Separating score generation from coreset construction enables multiple fixed-budget coresets to be constructed without retraining a scoring model.