Skip to content

Add refill strategy Implementation and some evaluation results #816

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.

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
wants to merge 2 commits into
base: main
Choose a base branch
from

Conversation

tengxiao1
Copy link

This PR implements the refill strategy in the training pipeline and documents the resulting performance improvements.

The refill strategy fills incomplete training batches with unfiltered prompts to maintain batch size and maximize data usage efficiency.

The comparison below is based on the same 3000 training steps on GRPO.

Benchmark Before Refill After Refill Δ (Change)
BBH 65.20 67.20 +2.00
MMLU 73.38 76.05 +2.67
PopQA 21.16 20.64 -0.52
ifeval 83.73 85.95 +2.22
GPQA 33.92 33.49 -0.43
AIME 7.78 11.11 +3.33
ZebraLogic 9.80 10.20 +0.40
MATH500 75.20 76.60 +1.40
GSM8K 88.02 89.84 +1.82
MATH 73.14 74.00 +0.86
AGI (average) 68.55

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant