Temazepam is a widely used insomnia drug, which normalizes sleep architecture and enhances slow-wave sleep. Our study aimed to showcase two methods of explaining a sleep-EEG classification model, in the input and representation space. The goal was to evaluate the effectiveness of these explanations in studying the influence of temazepam on sleep architecture. We fine-tuned BErt-inspired Neural Data Representations (BENDR) encoder pretrained on the TUH EEG Corpus dataset to sleep stage classification. The obtained (sleepBENDR) model is explained using Class Activation Mapping (CAM), which indicates the signal fragments relevant for classification. The representation structure developed in sleepBENDR is investigated by t-SNE. We used sleep-EEG of 10 subjects with mild sleep problems, each recorded in two conditions: placebo and after temazepam intake, to fine-tune sleepBENDR. Using a confusion matrix, we verified that the model performed on the state-of-the-art level in both cases. Qualitatively, the obtained hypnograms were similar to those prepared by the experienced scorers. The examined CAM explanations were consistent with American Academy of Sleep Medicine (AASM) scoring guidelines for sleep stages, as they often highlighted graphoelements typical for the predicted sleep stage. The t-SNE analysis indicates that the topology of the latent representation of sleepBENDR after temazepam intake is more consistent with normal sleep transitions than after placebo. The results demonstrate that explanations at the input space level allow us to validate the model against AASM recommendations. On the other hand, visualizations of the latent space reveal the effects of temazepam are consistent with prior knowledge about it.