The implementation of "Reinforced Preference Optimization for Recommendation" (ReRe).

File Description

• sft.py: the SFT code
• rere.py: the ReRe code
• rere_trainer.py: the GRPO trainer tailored for recommendation

The training instructions can be seen in rere.sh and train.sh, while the evaluation instructions are in evaluate.sh.

Quickstart

• Create a virtual Python environment.

conda create -n ReRe

• Install required packages.

pip install -r requirements.txt

• Execute the ReRe the training bash.

bash rere.sh

• Run the evaluation bash.

bash evaluation.sh