In real-world scenarios, the data collected often contains noise and biases. To ensure the performance of model training, it is essential to assess the quality of the data and perform appropriate selection. However, the data cleaning process cannot fully eliminate all low-value data points. Existing approaches, such as leave-one-out (LOO) and Shapley values, have been used to assess the importance of data. However, the LOO method has been shown to be unstable, while the computational complexity of the Shapley value method is excessively high. To balance usability with computational efficiency, this study proposes a gradient-based data value evaluation method, which computes the value of all data points more quickly than the Shapley value method. Experimental results demonstrate that the optimized gradient evaluation method effectively identifies mislabeled and noisy data, and the removal of low-value data points significantly improves the model’s predictive accuracy.