This work relies upon raw data to present the time-frequency consistency (TF-C) evaluation for human activity recognition (HAR). The original paper utilized data for this task in the pretext stage but did not explore its application in the downstream task. An application with a modified TF-C architecture uses HAR data on the downstream task, reporting an accuracy of 64.08%. We propose three experiments. First, we reproduce the original experiment with the epilepsy dataset, comparing the results with the reported ones. Second, we make a performance comparison test using different percentages of data from 0.1% to 100% and report the corresponding accuracy. Finally, we compare the results with supervised Convolutional Neural Networks and the supervised TF-C. This work demonstrates the feasibility of utilizing TF-C to perform HAR as downstream task, achieving an accuracy of 96% utilizing all data of the training dataset in fine-tuning. Even with just 42 samples of the training dataset, the model achieved an accuracy of 85% and to obtain an accuracy greater than 90% it is only necessary 126 train samples.