Reduce penalty weight

[steffencruz]

The current penalty weight of 1.0 effectively limits the max. score to 0.8 because there is always overlap between the challenge and completion texts. This was desirable at initial launch because we wanted to disincentivize echo mining, however such strategies are not viable at all any more.

Also, we don't want 'gaming' the penalty model to become profitable at the expense of producing high quality reference answers.

My recommendation is that we reduce the penalty to 0.5 for qa and summarization

[steffencruz]

Closed in #115