nathanwei05/cs2881r-dobby-qwen2.5-3b-final-combined-grpo100-step25
Harvard CS 2881R (AI Safety, Fall 2026), Assignment 1, stage 3 (combined). LoRA GRPO
(rank 16, LR 2e-5, beta 0.04, 100 updates, checkpoint 25 selected on a 150-prompt dev set)
from axel-sdq/cs2881r-dobby-qwen2.5-3b-rlaif-grpo100-step25 at revision ba9705301d86619447f480121c06133039a496d3, which is itself SFT +
RLAIF from Qwen2.5-3B-Instruct. Reward: persona x quality / 16 from a DeepSeek judge + verifier correctness - 0.5 x degeneration penalty (loop or missing EOS), weights 1 / 1 / 0.5.
The merged model is at the repository root; the LoRA adapter is under adapter/.
Training code, data, worklog and results: https://github.com/harvard-cs2881f26/hw1-soderquist-wei