These datasets would be structured and presented in a way that is accessible for students or scholars interested in machine learning. Students could use them to learn how to create applications that utilize machine learning, and scholars could use them in research/studies.
These datasets come in many different formats and topics. The oldest datasets in the repository date back to the late 80s, and there are some datasets that are from 2013.
Hilary Mason (a data scientist at bit.ly and speaker on the topic of Machine Learning) put together a list of "Research Quality" datasets in a bit.ly bundle specifically for this reason
It has some interesting datasets from a variety of different fields ripe for Machine Learning research
The Weka machine learning toolkit (in Java) introduced the text-based ARFF format (Attribute-Relation File Format) as a proposed standard format for machine learning datasets. Its website has a list of ARFF datasets. Some of them are discussed in the excellent Weka book that is now in its third edition: Witten, Frank, & Hall (2011): „Data mining. Practical machine learning tools and techniques“, Morgan Kaufmann Publishers, ISBN: 978-0-12-374856-0.
If you are using R, there is the R Datasets Package. In this way, the data is prepared and ready for analyis in R.
There are also examples of the R code, which includes how to use the library. See here for the canonical iris data set.
Other languages will also have sample datasets. In python, the scikit library is popular for machine learning and predictive analytics.
microformats are the most abundant semantic technology deployed on the web today, aka they are the biggest dataset for machine learning. as long as the html document is marked up properly, they fit all of your needs and then some. you can find sites that are deploying them in the "wild" at the wiki:
and if you do a quick web search, you'll find plenty of open source microformat parsers and scrapers, code snippets, and libs, to let your gorge on the semantic markup yumminess.
There is also MLcomp, whose concept of transparent comparison of ML algorithms is very attractive to me. Unfortunately the website is a bit dormant nowadays.
Note that most of the time datasets are grouped by machine learning task, e.g. for recommendation systems:
You might find this curated list (by some well known names in the field) interesting.