search

PygmalionAI / data-toolbox

Our data munging code.
GNU Affero General Public License v3.0
34 stars 9 forks source link

Curtail curtailing harmful outputs #19

Closed TearGosling closed 1 year ago

TearGosling commented 1 year ago

Recent versions of Metharme-13B have been reported to have issues with refusals to do certain types of roleplay. While this could be due to many factors, one possible factor could be directions in the system prompt to "avoid harmful output." Initially, we thought that this would only happen if you specifically prompted it to, but it might be possible that either the model is still too small to be able to keep it contained or some other factor makes it so that it affects the model's outputs in general.

We can take a step to avoid this by removing directions asking the model to restrict harmful output. In addition, I've added a prompt to the roleplay task that mimics SillyTavern's roleplay system prompt.

TearGosling commented 1 year ago

Automatically merging pull request due to 11b's absence.