Skip to content

feat: Default loss to CISPO #498

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
bradhilton merged 1 commit into from
Jan 2, 2026
+10 −3
Merged

feat: Default loss to CISPO #498

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
13 changes: 10 additions & 3 deletions src/art/loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,8 +65,15 @@ def loss_fn(
prob_ratio = (prob_ratio + sequence_prob_ratio) / 2
elif importance_sampling_level == "geometric_average":
prob_ratio = (prob_ratio**0.5) * (sequence_prob_ratio**0.5)
epsilon = experimental_config.get("epsilon", 0.2)
epsilon_high = experimental_config.get("epsilon_high", epsilon)
ppo = experimental_config.get("ppo", False)
if ppo:
epsilon_default = 0.2
epsilon_high_default = None
else:
epsilon_default = 1.0
epsilon_high_default = 4.0
epsilon = experimental_config.get("epsilon", epsilon_default)
epsilon_high = experimental_config.get("epsilon_high", epsilon_high_default)
if epsilon_high is None:
epsilon_high = epsilon
if max_negative_advantage_importance_sampling_weight := experimental_config.get(
Expand All @@ -83,7 +90,7 @@ def loss_fn(
)
if tau := experimental_config.get("kimi_k2_tau", None):
advantages -= tau * logprob_diff.detach()
if experimental_config.get("ppo", True):
if ppo:
policy_loss = -torch.min(
prob_ratio * advantages,
torch.clip(prob_ratio, 1 - epsilon, 1 + epsilon_high) * advantages,
Expand Down