Skip to content

[RL] Fix loss: use global token normalization instead of per-example #2376

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
AlienKevin merged 3 commits into from
Jan 19, 2026
Merged

[RL] Fix loss: use global token normalization instead of per-example #2376

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
5254938
Fix DAPO loss: use global token normalization instead of per-example
AlienKevin Jan 17, 2026
0669d3c
Merge branch 'main' into kevin/rl-fix-regression
AlienKevin Jan 19, 2026
f707541
Merge branch 'main' into kevin/rl-fix-regression
AlienKevin Jan 19, 2026
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
8 changes: 5 additions & 3 deletions lib/marin/src/marin/rl/rl_losses.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -235,9 +235,11 @@ def compute_dapo_loss(
loss_objective: jax.Array,
loss_masks: jax.Array,
) -> jax.Array:
"""Compute DAPO-like loss (per-example normalization)."""
# Use per-example normalization (averaging the per-example means)
return -1 * jnp.mean(jnp.sum(loss_objective * loss_masks, axis=1) / jnp.sum(loss_masks, axis=1))
"""Compute DAPO-like loss (global token normalization).

Divides by total tokens across all examples in the batch, not per-example.
"""
return -1 * jnp.mean(jnp.sum(loss_objective * loss_masks, axis=1) / jnp.sum(loss_masks))


def compute_grpo_loss(
Expand Down