Good training data's partly hard to come by because there's often reasonably poor inter-annotator agreement on sentiment datasets -- that is to say, humans disagree a lot in how we interpret a phrase. What reads like sarcasm to you might read like genuine enthusiasm to another.

It's pretty easy to load up a set of data into a crowdsourcing tool and use microtasks to rate it, but my experiences doing so weren't superb (even restricting to native English speakers alone).

A better source of data is starred reviews where you have the star rating and the review itself -- these come free with a sentiment rating, although plenty of caveats around normalization. There are lots of places with review systems like this and some (like Yelp) even make the data available: https://www.yelp.com/academic_dataset