This repository is designed for implementing LLM-SRec.
In this paper, we first demonstrate through a series of experiments that existing LLM4Rec models do not fully capture sequential information both during training and inference. Then, we propose a simple yet effective LLM-based sequential recommender, called LLM-SRec, a method that enhances the integration of sequential information into LLMs by distilling the user representations extracted from a pre-trained CF-SRec model into LLMs.
- We use LLaMA-3.2-3b-instruct.
conda create -n [env name] pip
conda activate [env name]
pip install -r requirements.txt
The data (Amazon 2023) is automatically downloaded when using the SASRec training code provided below.
cd SeqRec/sasrec
python main.py --device 0 --dataset Industrial_and_Scientific
We have also implemented LLM-SRec on the Gaudi-v2 environment (Note that --nn_parameter must be used for training models on Gaudi-v2):
python main.py --device hpu --dataset Industrial_and_Scientific --nn_parameter
The model saves when the best validation score is reached during training and performs inference on the test set.
python main.py --device 0 --train --rec_pre_trained_data Industrial_and_Scientific --save_dir model_train --batch_size 20
For Gaudi-v2:
python main.py --device hpu --train --rec_pre_trained_data Industrial_and_Scientific --save_dir model_train --batch_size 20 --nn_parameter