You're assuming the training data is a true representation of language usage, which it isn't. In the "gray" or "overcast" case, a single weather website might

You're assuming the training data is a true representation of language usage, which it isn't. In the "gray" or "overcast" case, a single weather website might already skew the count of these words.