Word representation datasets for document-level tasks?

First of all, have you actually tried using this data in document-level tasks? My intuition is that it will probably make things better, even if it was generated with more non-local properties in mind.

The key question then is how to aggregate individual word representations into a document representation you can use in your feature vector. If this is your main concern I suggest you use a probabilistic topic model (such as LDA) to generate topic features. Dan Ramage suggests that making the dot product between two documents being 0.8*(tf-idf dot product) + 0.2*(dot product of document/topic allocations, normalized) is a sweet spot that's usually hard to beat.

This paper uses distributed representations in document classification task. Although the method they used to induce distributed representations is not specifically intended for document classification, but these kinds of model can be modified to induced distributed representations for document classification.

In fact word representations are being used in supervised learning algorithms to generalize the parameters of unseen related words while testing in test set, these kinds of representations can improve the performance if they are appended as extra features in already existing supervised learning setup.