The paper reports an improvement of an unsupervised classification system for learner peculiarities of flat finishing motion of an iron file in skill training. The system classifies and visualizes peculiarities of learners' tool motion effectively by using a torus type Self-Organizing Maps (SOM). An automatic clustering based on codebook density of SOM helps skill trainers to grasp learners' peculiarities distribution easily. In this paper, we focus on the classification improvement of the SOM on the viewpoint of the clustering. Classification performance is improved by comparing different learning schedules of the SOM. Effectiveness of the improvement is evaluated with measured data of an expert and sixteen learners.