lm_human_preferences/rewards.py (1 line):
	- line 13: # TODO: combine this with TrainedRewardModel


lm_human_preferences/train_policy.py (1 line):
	- line 199: # NOTE: must line up with stats created in self.loss (TODO: better solution?)


lm_human_preferences/language/model.py (1 line):
	- line 367: # TODO: refactor this, perhaps move to Policy


lm_human_preferences/train_reward.py (1 line):
	- line 80: # TODO: download on just one rank?  then do: labels = utils.mpi_bcast_tensor_dict(labels, comm=comm)