negative cost with negative weights

[tomsbergmanis]

The documentation says that word weights can be real-valued. Real numbers can be negative. This is convenient because outputs of models one could use for scoring often output log-likelihoods, which are negative numbers.

However, using negative weights results in a negative cost value, which, if minimized, maximizes the error:

• If this is a mistake in the documentation, could we update the documentation to say "non-negative numbers"?
• If not, could we update the documentation with the steps to be taken to avoid divergence with negative weights?

Cheers,
Toms