fix(opd): score teacher logprobs at rollout temperature, not 0#2085
Open
EazyReal wants to merge 1 commit into
Open
fix(opd): score teacher logprobs at rollout temperature, not 0#2085EazyReal wants to merge 1 commit into
EazyReal wants to merge 1 commit into
Conversation
The on-policy-distillation teacher reward_func scored teacher log-probs via SGLang with a hardcoded `temperature: 0`. SGLang computes input_token_logprobs WITH temperature scaling (compute_temp_top_p_normalized_logprobs), and the student log-probs are temperature-scaled by rollout_temperature (get_responses). So when rollout_temperature != 1 the OPD reverse-KL (student - teacher) compares log-probs at different effective temperatures and is biased. Score the teacher at rollout_temperature so both sides of the KL match. No change at the default rollout_temperature=1.0. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Problem
The on-policy-distillation teacher
reward_func(slime/rollout/on_policy_distillation.py) scores teacher log-probs via SGLang with a hardcodedtemperature: 0, then uses them for the OPD reverse-KLstudent - teacher.SGLang computes
input_token_logprobswith temperature scaling (compute_temp_top_p_normalized_logprobs), and the student log-probs are temperature-scaled byrollout_temperature(get_responses). So whenrollout_temperature != 1, the two sides of the OPD KL are at different effective temperatures → the distillation signal is biased.Fix
Score the teacher at
rollout_temperature(same as the student), so the KL is between same-temperature distributions. No change at the defaultrollout_temperature=1.0. (miles counterpart: radixark/miles#1345.)🤖 Generated with Claude Code