Training the default perceptron tagger with bigger corpus

The default pickle for the PerceptronTagger is nice but it's a little too small for realistic usage:

>>> from nltk import PerceptronTagger
>>> len(PerceptronTagger(load=True).tagdict)
1549

Would it be possible to retrain a model on something like the full Penn TreeBank or BNC? Is the model then releasable on NLTK?