My first thought would be arXiv, where you can find 973,256 full-text papers.
You may not bulk-download them directly from arXiv.org, but other methods are available. In particular, you can download every PDF hosted by arXiv via Amazon S3, by paying a bandwidth charge of USD 0.12 per gigabyte. See these instructions. Metadata in various formats is also available.
Contrary to DTCLib's assertion that "nobody could provide you with such a data set without violating copyright laws", it is perfectly legal for arXiv to do this, since each author submitting a paper to the repository gives arXiv the right to redistribute the paper to anyone. Note, however, that does not in general give you the right to redistribute them further. (Some papers on arXiv are released under Creative Commons licenses, or even are in the public domain, but most are not.)