Skip to content

fix：fix bugs in cosine reward of GRPO #3358

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Mar 3, 2025
Merged

fix：fix bugs in cosine reward of GRPO #3358

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8fecfbe
Update grpo_trainer.py
youyc22 Mar 3, 2025
9554a20
Update orm.py
youyc22 Mar 3, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion swift/plugin/orm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -296,12 +296,14 @@ def __call__(self, completions, **kwargs) -> List[float]:
class CosineReward(ORM):
# https://arxiv.org/abs/2502.03373
def __init__(self,
tokenizer=None,
cosine_min_len_value_wrong: float = 0.0,
cosine_max_len_value_wrong: float = -0.5,
cosine_min_len_value_correct: float = 1.0,
cosine_max_len_value_correct: float = 0.5,
cosine_max_len: int = 1000,
accuracy_orm=None):
self.tokenizer = tokenizer
self.min_len_value_wrong = cosine_min_len_value_wrong
self.max_len_value_wrong = cosine_max_len_value_wrong
self.min_len_value_correct = cosine_min_len_value_correct
Expand All @@ -326,7 +328,7 @@ def __call__(self, completions, solution, **kwargs) -> List[float]:
else:
min_value = self.min_len_value_wrong
max_value = self.max_len_value_wrong
gen_len = len(content)
gen_len = len(self.tokenizer.encode(content))
reward = self.cosfn(gen_len, self.max_len, min_value, max_value)
rewards.append(reward)
return rewards
Expand Down
3 changes: 3 additions & 0 deletions swift/trainers/rlhf_trainer/grpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,6 +71,7 @@ def __init__(self,
self.args = args
self.queue = Queue()
self.processing_class = kwargs.get('template').tokenizer
self.tokenizer = kwargs.get('template').tokenizer if hasattr(kwargs.get('template'), 'tokenizer') else None
if not isinstance(reward_funcs, list):
reward_funcs = [reward_funcs]

Expand All @@ -83,6 +84,8 @@ def __init__(self,
key: getattr(args, key)
for key in reward_func_args if key not in ['self', 'args', 'kwargs'] and hasattr(args, key)
}
if reward_func_class.__name__ == 'CosineReward' and 'tokenizer' in reward_func_args:
reward_func_kwargs['tokenizer'] = self.tokenizer
reward_funcs[i] = reward_func_class(**reward_func_kwargs)
elif not callable(reward_func):
raise ValueError(f'reward_function {reward_func} is not implemented in swift.llm.plugin')
Expand Down