Evaluating Sample Utility for Data Selection by Mimicking Model Weights