feat: add RLSD (Self-Distilled RLVR) credit assignment#333
Open
morgendave wants to merge 1 commit intomainfrom
Open
feat: add RLSD (Self-Distilled RLVR) credit assignment#333morgendave wants to merge 1 commit intomainfrom
morgendave wants to merge 1 commit intomainfrom
Conversation
Adds composable per-token credit weighting from self-distillation evidence ratios (arXiv 2604.03128). RLSD modulates the per-token advantage magnitude using the teacher-student log-prob ratio: w_t = clip(exp(sign(A) · (log P_teacher - log P_student)), 1±ε_w) Key design: RLSD is a weight that composes with any existing loss (GRPO, DAPO, CISPO, etc.) — just like TIS. No new loss function. Files: - training/utils/rl/rlsd.py: RLSDConfig + compute_rlsd_weights() - training/utils/rl/common.py: SampleContext.rlsd_weight + run_loss_loop plumbing - training/utils/rl/grpo.py: multiply rlsd_weight into surrogate loss - training/utils/rl/__init__.py: export RLSDConfig
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds composable per-token credit weighting from self-distillation evidence ratios (arXiv 2604.03128). RLSD modulates the per-token advantage magnitude using the teacher-student log-prob ratio:
w_t = clip(exp(sign(A) · (log P_teacher - log P_student)), 1±ε_w)
Key design: RLSD is a weight that composes with any existing loss (GRPO, DAPO, CISPO, etc.) — just like TIS. No new loss function.
Files: