Datasets

Labeled data that could be used for captcha tasks:

https://www.cs.toronto.edu/~kriz/cifar.html - labeled pictures organized in categories, e.g. cat, bird

http://deeplearning.net/datasets/ - contains links to different data sets previously used for machine learning e.g. http://yann.lecun.com/exdb/mnist/ - collection of labeled handwritten digits

Note: Didn't find any information, whether they have copyright on them or not, e.g. if it' allowed to use them for more than training networks