This PR decompose out the function doc for the multi-turn categories from the dataset file itself. This helps view the changes (otherwise git diff would load forever), and we programmatically propagate the function docs for individual test entries, reducing the file sizes.
This PR decompose out the function doc for the multi-turn categories from the dataset file itself. This helps view the changes (otherwise
git diffwould load forever), and we programmatically propagate the function docs for individual test entries, reducing the file sizes.This PR will not affect the leaderboard score.