TensorFlow is the machine learning codebase, but typically how do machine learning research teams manage their training sets, dataset metadata and collaboration on these large datasets?
IDSIA, affiliated with Juergen Schmidhuber and many other leading ML researchers, has released Sacred, "a tool to help you configure, organize, log and reproduce experiments." https://github.com/IDSIA/sacred
MILA, affiliated with Yoshua Bengio and the Theano project, offers fuel, "a data pipeline framework for machine learning": https://github.com/mila-udem/fuel
It requires a fair amount of set-up, but works surprisingly well once there is a core team and problems established.
We are building mldb.ai to help bring the data and the algorithms for ML together in a less ad-hoc manner and to help move things out of research and into prod once they are ready. Many of the hosted ML solutions (Azure ML, Amazon ML, Google Data Lab, etc) and other toolkits (eg Graphlab) are working on similar ML workflow and organizational structure problems.