For example, I don't know where to look for the sentences you speak about, so maybe someone could give me some clue about it...
To train the model you need to feed it with a huge volume of text. The text should be representative of the kind of text you will type. For example is you use a Wikipedia corpus, the keyboard will be very uncooperative if you try to type informal text that would look unnatural in a Wikipedia article. Building language files is not just a matter of pouring random text in the build tool or you will end up with a high error rate. I recommend using a lot of text (my French corpus is over 40 million words, and in some cases this is not enough), and using different kind of documents: articles (new / wikipedia), e-mail, IRC and chat logs ...