The choice of training population (TP) is an important criterion in genomic selection. In this study, 200 F5 wheat lines were selected based on a k-means clustering method, partitioned into three cluster (K-1, K-2, and K-3), to represent a portion of the TP. A supplementary 300 F5 lines selected based on pedigree, and 45 parental lines were also included to make a TP with a total size of 545 lines. Our objective was to assess the prediction accuracies for three traits associated with Fusarium Head Blight in three different scenarios: S1) using entire TP to predict itself (n = 545), S2) using entire TP to predict each cluster and S3) using each cluster with parental lines to predict itself (n = 140 in K-1, 91 in K-2, and 104 in K-3). The leave-one-out-cross-validation (loocv) method in RR-BLUP was used to assess the predictive abilities in the three scenarios. The TP was evaluated for disease severity (SEV), visually scabby kernels (VSK) and micro-test weight (MTWT) in St Paul, Minnesota in 2018. Prediction accuracies (rMP) ranged from 0.30 in VSK to 0.39 in MTWT in S1. In S2, rMP was highest when loocv was performed on K-3 with values ranging from 0.4 in SEV to 0.55 in MTWT while K-2 performed best in S3 with rMP values that ranged from 0.38 in SEV to 0.58 in MTWT. Since rMP in S1 and S2 did not significantly outperform that of S3, we concluded that a TP of size 91 – 140 is as efficient as a TP of size 545.