📋 A README.md for code accompanying our paper RPP

Reinforced Prompt Personalization for Recommendation with Large Language Models

1. Prepare the environment

Prepare the environment to evaluate on ChatGPT and LLaMa2-7b-chat

cd RPP
pip install -r requirements.txt

Prepare the environment to evaluate on Alpaca

pip install -r requirements_A.txt

2. Prepare the pre-trained LLMs

Insert the API of gpt-3.5-turbo to openai_api.yaml to test the performance of ChatGPT

Download the pre-trained Bert model to transform prompts into embedding(https://huggingface.co/bert-base-chinese)

Download the pre-trained LLaMa2-7B-chat model (https://ai.meta.com/llama/)

Download the pre-trained huggingface model of LLaMA2-7B to finetune Alpaca (https://huggingface.co/meta-llama/Llama-2-7b-hf)

2. Prepare the checkpoints of RPP and finetuned Alpaca

Put the checkpoints of RPP to the dir path saved/ and the checkpoints of finetuned Alpaca to the dir path model/

Training

We use the NVIDIA A100 to conduct all our experiments on LLMs:

For the dataset of ML-1M and Lastfm, the initial interaction history length (few-shot) ini_len: 1

For the dataset of Games, the initial interaction history length (few-shot) ini_len: 5

To train the frozen LLM of LLaMa2-7B-chat:

selected_user_suffix: train
CUDA_VISIBLE_DEVICES=[] torchrun --nproc_per_node 1 --master_port=[] train.py

To evaluate the frozen LLM of LLaMa2-7B-chat:

selected_user_suffix: test
CUDA_VISIBLE_DEVICES=[] torchrun --nproc_per_node 1 --master_port=[] evaluate.py