Logging win rate used to depend on config.env.eval_metric == "win_rate" however in certain envs we don't want the win rate to be the eval_metric as it is too sparse to optimize for, but we do still want to see the win rate.
This is a stop gap solution for now, we should spend time finding out how we can cleanly log all env metrics regardless of name.
What?
Logging win rate used to depend on
config.env.eval_metric == "win_rate"however in certain envs we don't want the win rate to be the eval_metric as it is too sparse to optimize for, but we do still want to see the win rate.This is a stop gap solution for now, we should spend time finding out how we can cleanly log all env metrics regardless of name.