fix(sc): gate reference logprobs on KL penalty - #3742

Merged

yfw merged 2 commits into

NVIDIA-NeMo:mainfrom

jinglinglingling:linglinj/v2-zero-kl-reference-logprobs

Sep 1, 2026

Commits on Aug 29, 2026

fix(sc): gate reference logprobs on KL penalty

jinglinglingling
authored and
Linglin Jing
committed
dffc7f2
fix(sc): align train fields with logprob gating

jinglinglingling
authored and
Linglin Jing
committed
ff0df3f