Interesting - seems it's mostly just trying to fit one experts idea of what valence means, so we'd have to ask that dude what features contribute to valence.

But... aside from the fact that it's relatively old and predates a lot of the better language models we have today, I think we can also conclude it's likely not looking at lyrical content by the inclusion of both Hey ya and Pumped Up Kicks at the top of the valence scores.