In this paper, we construct facial expression benchmark data of 100 persons using Kinect face tracking application and study the stability of the benchmark data in terms of clustering. Kinect with its Software Development Kit applications has enabled low-cost constructions of various benchmark data on humans. We devised multi-lingual instruction sheets on 25 expressions, collected data from 115 persons, and carefully inspected and labeled the outcome to construct the data. The benchmark data consist of 263,106 instances, each of which includes 6 animation units, 11 shape units, and an image file all provided by the application. Out of the 263,106 instances, we labeled 62,500 of them as 1 of the 25 expressions and investigated their clustering stabilities to the 17 features. We show that the most frequently used clustering algorithm: k-means achieves the average normal mutual information about 0.92 as an evidence of the stability of our facial expression data.