Skip to content

[MLPerf 6.1][GRPO] Update Qwen3.5 reference assets #904

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
jepio wants to merge 1 commit into
base: master
Choose a base branch
from
Open

[MLPerf 6.1][GRPO] Update Qwen3.5 reference assets #904

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
2 changes: 1 addition & 1 deletion llm_post_training/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -659,13 +659,13 @@ The recipe uses token-level GRPO with:
- a leave-one-out baseline;
- no reward shaping;
- no reference-policy KL penalty (`reference_policy_kl_penalty: 0`);
- asymmetric PPO ratio clipping configured as `0.2` and `0.28`;
- asynchronous collection, with the checked-in RCPs using a maximum trajectory
age of one training step;
- sequence-mask truncated importance sampling (`seq-mask-tis`) with lower and
upper bounds `0.999` and `1.002`;
- importance-sampling correction enabled;
- `force_on_policy_ratio: true`;
- asymmetric PPO ratio clipping is not used due to `force_on_policy_ratio: true` (any configured values are ignored)

@jepio jepio Sep 18, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@susanbao capturing the note based on our discussion in the WG meeting yesterday

- reference-policy log-probability calculation skipped;
- sequence-level log-probability error masking with threshold `2.0`; and
- overlong-response filtering enabled.
Expand Down
Loading