This repo contains the code for the paper, KASER: Knowledge-Aligned Student Error Simulator for Open-Ended Coding Tasks, by Zhangqi Duan, Nigel Fernandez, and Andrew Lan, published at ACL 2026. We will release the cleaned code soon.
The following will train, test, and evaluate the model. We use Qwen2.5-Coder-7B-Instruct as the base model.
python sft.py
python grpo.py --data_path data/grpo_student_knowledge_notrunc.pkl --with_knowledge