In a typical fusion experiment, the plasma can have several possible confinement modes. At the tokamak a configuration variable, aside from the low (L) and high (H) confinement modes, an additional mode, dithering (D), is frequently observed. Developing methods that automatically detect these modes is considered to be important for future tokamak operation. Previous work (Matos et al 2020 Nucl. Fusion 60 036022) with deep learning methods, particularly convolutional long short-term memory networks (conv-LSTMs), indicates that they are a suitable approach. Nevertheless, those models are sensitive to noise in the temporal alignment of labels, and that model in particular is limited to making individual decisions taking into account only the input data at a given timestep and the past data, represented in its hidden state. In this work, we propose an architecture for a sequence-to-sequence neural network model with attention which solves both of those issues. Using a carefully calibrated dataset, we compare the performance of a conv-LSTM with that of our proposed sequence-to-sequence model, and show two results: one, that the conv-LSTM can be improved upon with new data; two, that the sequence-to-sequence model can improve the results even further, achieving excellent scores on both train and test data.