Fix key error in GRPOTrainer

[le-big-mac]

Fixes #1807

Currently, when using GRPOTrainer, a KeyError occurs in the loss computation due to a missing [mode] key. This PR adds this mode key back.

[danielhanchen]

Thanks! Oh this is for the nightly release of TRL right?
Would it be possible to first check whether self._metrics["completion_length"] exists, and or self._metrics["train"] / "eval" exists to allow older versions of TRL to work? Also it looks like I might need to update the notebook logging

[le-big-mac]

I've added a check for "train" as a key in self._metrics, from looking at the history of TRL's GRPOTrainer this should be enough to distinguish between the versions. It used to be self._metrics = defaultdict(list), and now it's self._metrics = {"train: defaultdict(list), "eval": defaultdict(list)}, and I can't see evidence of a "train" key ever being used in the old version.

[danielhanchen]

Very good work thanks!