Closed davidbrandfonbrener closed 7 months ago
This pull request fixes a bug where we were selecting the wrong index for features to use in run_diversity.
Before we were just selecting the features associated with the last token, but the last token was often a padding token yielding incorrect features.
Now we correctly select the feature from the index corresponding to the end of the actual generated text.
This pull request fixes a bug where we were selecting the wrong index for features to use in run_diversity.
Before we were just selecting the features associated with the last token, but the last token was often a padding token yielding incorrect features.
Now we correctly select the feature from the index corresponding to the end of the actual generated text.