The stochastic block model is a classical cluster-exhibiting random graph model. In its simplest form there are two equal-sized clusters with intra-cluster edge probability p and inter-cluster edge probability q < p. We study a labelled version of this model, where in addition to the graph structure, for a small fraction of the nodes the true cluster labels are also revealed. Here, for the case of two clusters we show that for any given node, local information (labelled nodes in a small neighbourhood) suffices to predict to optimal accuracy. In the regime where there are large number of clusters, we show that reconstruction of underlying clusters is possible even below the conjectured algorithmic threshold in the standard (unlabelled) model. We also report preliminary simulation results.