Skip to content

fix(rewards): make DAPO zero-penalty region inclusive so cache_length=0 no longer divides by zero #627

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

Open
Linxiushen wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(rewards): make DAPO zero-penalty region inclusive so cache_length=0 no longer divides by zero #627

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
26 changes: 26 additions & 0 deletions tests/algorithm/test_dapo_algorithm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,6 +108,32 @@ def test_symmetric_accuracy(self, mock_compute_score):
self.assertEqual(good["accuracy"], 1.0)
self.assertEqual(bad["accuracy"], -1.0)

def test_overlong_penalty_piecewise(self):
fn = MathDAPORewardFn(
enable_overlong_penalty=True,
penalty_factor=1.0,
max_response_length=100,
cache_length=20,
)
# DAPO paper (Sec. 2.4): zero until `max_response_length - cache_length`,
# then linear, then -penalty_factor beyond `max_response_length`.
self.assertEqual(fn.compute_overlong_penalty(torch.zeros(79)), 0.0)
self.assertEqual(fn.compute_overlong_penalty(torch.zeros(80)), 0.0)
self.assertAlmostEqual(fn.compute_overlong_penalty(torch.zeros(90)), -0.5)
self.assertAlmostEqual(fn.compute_overlong_penalty(torch.zeros(100)), -1.0)
self.assertEqual(fn.compute_overlong_penalty(torch.zeros(101)), -1.0)

def test_overlong_penalty_without_soft_window(self):
fn = MathDAPORewardFn(
enable_overlong_penalty=True,
penalty_factor=1.0,
max_response_length=100,
cache_length=0,
)
self.assertEqual(fn.compute_overlong_penalty(torch.zeros(99)), 0.0)
self.assertEqual(fn.compute_overlong_penalty(torch.zeros(100)), 0.0)
self.assertEqual(fn.compute_overlong_penalty(torch.zeros(101)), -1.0)


if __name__ == "__main__":
unittest.main()
2 changes: 1 addition & 1 deletion trinity/common/rewards/dapo_reward.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -89,7 +89,7 @@ def compute_overlong_penalty(self, response_token):
response_len = len(response_token)
expected_len = self.max_response_length - self.cache_length

if response_len < expected_len:
if response_len <= expected_len:
return 0.0
elif response_len > self.max_response_length:
return -self.penalty_factor
Expand Down