Reproduce “GRPO Training Dynamics for Small Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rajat Ghosh, Vaishnavi Bhargava, Henry Wong, Aryan Singhal, Debojyoti Dutta
The paper studies GRPO in small language models across math, MCQ, and code tasks on one 8×A100 node. It scores groups of answers, updates by relative reward, and avoids a critic. On GSM8K, DeepSeek-7B leads reward, while 1.5B DeepSeek-R1-Qwen leads accuracy, especially at G=16. Diagnostics guide LoRA tuning; APPS-Intro pass@1 reaches 11.83% versus 6.28% baseline. MCQ gains remain limited.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.