Add alpaca reverse augmentation possibility #2342
Conversation
|
❌ pre-commit failed. |
|
Eval dataset is still missing |
|
One thing that is not quite clear to me is if we should the reverse augmentation samples ADDITIONALLY to the train set |
Yes, I think forward + reverse could both be part of the training set. Ideally the same examples of the alpaca set should be used for evaluation of both forward/reverse (in order to avoid leaking of eval-data of the other dataset into the train set). |
For the smaller models it is necessary to specify |
|
Added the eval set and also the option to keep the unreversed data in the training set. |
CloseChoice
requested review from
theblackcat102,
sanagno,
dvruette,
andreaskoepf and
yk
as code owners
| super().__init__() | ||
| self.data = data | ||
| if mode not in ["sft", "rl"]: |
There was a problem hiding this comment.
could be tuple instead of list
| class Alpaca(AlpacaBase): | ||
| def __init__(self, mode: str = "sft", cache_dir: str = None) -> None: | ||
| super().__init__(dataset_name="yahma/alpaca-cleaned", mode=mode, cache_dir=cache_dir) | ||
| class AlpacaDataset(AlpacaBaseDataset): |
There was a problem hiding this comment.
If these classes don't add any functionality they could be removed (i.e. base class be used directly).
| manual_seed: int = 287631038922, | ||
| reverse_augmentation: bool = False, | ||
| keep_unreversed: bool = True, | ||
| ) -> tuple[AlpacaDataset, AlpacaDataset] | tuple[CodeAlpacaDataset, CodeAlpacaDataset]: |
There was a problem hiding this comment.
base class type annotation for factory function would be better imo
closes #2335
Alpaca with reverse augmentation can be run with:
I couldn't run this due to unscale gradient errors, so before this gets merged we should do a run to check whether this reversal really improves the loss.