For the training of the BP neural networks in CRT color conversion, some papers suggest using a uniformly distributed RGB training set model (URGB). However, this URGB model is single-directional. Therefore, when the number of the samples in a training set is under a certain amount, such as less than 51 2 (8 X 8 X 8), a URGB model may cause big prediction errors, especially in the backward conversion (XYZ to RGB). In this paper, we propose an improved training set model, with which a smaller training set can be drawn from a virtual URGB set. Our experimental results show that, an improved training set model can achieve a desired prediction accuracy in the whole CRT color space, even if the samples number in a training set is less than 512(8 X 8 X 8).