The idea is to use the dataset to validate a multivariate method for signal analysis and I would need something sharp and robust like finger tapping or else.
Thank you so much for your suggestion. I am afraid however, that the databases at OpenfMRI do not meet my criteria. Gorgolewski et al.'s database (ds000114) encompasses a motor paradigm, but involving just 10 subjects, and 10 subjects is very short for me because I need to split the sample in train and test (in order to validate an ANN - artificial neural network).