Sleep is considered to support the formation of skill memory. In juvenile but not adult song birds learning a tutor's song, a stronger initial deterioration of song performance over night-sleep predicts better song performance in the long run. This and similar observations have stimulated the view of sleep supporting skill formation during development in an unsupervised off-line learning process that, in the absence of external feedback, can initially also enhance inaccuracies in skill performance. Here we explored whether in children learning a motor sequence task, as in song-learning juvenile birds, changes across sleep after initial practice predict performance levels achieved in the long run. The task was a serial reaction time task (SRTT) where subjects had to press buttons which were lighted up in a repeating eight-element sequence as fast as possible. Twenty-five children (8-12 years) practised the task in the evening before nocturnal sleep which was recorded polysomnographically. Retrieval was tested on the following morning and again 1 week later after daily training on the SRTT. As expected, changes in response speed over the initial night of sleep were negatively correlated with final performance speed after the 1-week training. However, unlike in song birds, this correlation was driven by the baseline speed level achieved before sleep. Baseline-corrected changes in speed or variability over the initial sleep period did not predict final performance on the trained SRTT sequence, or on different sequences introduced to assess generalization of the trained behaviour. The lack of correlation between initial sleep-dependent changes and long-term performance might reflect that the children were too experienced for the simple SRTT, possibly also favouring ceiling effects in performance. A consistent association found between sleep spindle activity and explicit sequence knowledge alternatively suggests that the expected correlation was masked by explicit memory systems interacting with skill memory formation.