Skip to content

Lower default RL weight_decay from 0.01 to 0.001 for LoRA #5747

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
danielhanchen merged 1 commit into from
May 24, 2026
+3 −1
Merged

Lower default RL weight_decay from 0.01 to 0.001 for LoRA #5747

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
4 changes: 3 additions & 1 deletion unsloth/models/rl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1312,7 +1312,9 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
"logging_nan_inf_filter": False,
"per_device_train_batch_size": 4,
"gradient_accumulation_steps": 2,
"weight_decay": 0.01,
# LoRA decays A and B toward 0 so effective W = W_init + (alpha/r) * B @ A is pulled toward W_init, not 0 as in full FT.
# 0.001 keeps a small Frobenius prior |A|_F^2 + |B|_F^2 without measurably dragging the merged adapter back to base.
"weight_decay": 0.001,
"seed": 3407,
"optim": "adamw_8bit",
"learning_rate": 5e-05,
Expand Down
Loading