Code inplement of dual-graph sequential model and finetuning LLaMA2 of URLLM
./DG_Final/GM processed movie-game dataset
./DG_Final/AO processed art-office dataset
./DG_Final path of Dual Graph Sequence Modeling Model of URLLM
./DG_Final/DG_src/dataset/Entertainment-Education_Amazon user interaction in movie-game dataset
./DG_Final/DG_src/dataset/Entertainment-Education_AO user interaction in art-office dataset
./DG_Final/DG_src/dataset/item_prompt_GM item attribute in movie-game dataset(generated from GPT)
./DG_Final/DG_src/dataset/item_prompt_AO item attribute in art-official dataset(generated from GPT)
./llama2-SFT path of finetuning LLaMA2
the following are pipelines of running URLLM on movie-game dataset. The same as the art-office dataset.
python ./DG_Final/GM/jsonBuilder_attribute_graph_GPT.py to gain prompts for GPT to build attribute graph
python ./DG_Final/GM/testGPT35 to gain prompts for GPT to build attribute graph
run sh ./DG_Final/DG_src/train.sh to gain the similarity of movie-game dataset user to generate user representations like DGGM_final_test_x_fea.npy.
python ./DG_Final/GM/Final_train_contrasive_searcher.py to gain user retrieval matrix best_trte_XORY_DG_.npy
python ./DG_Final/GM/jsinBuilder_testing_rc.py utilizing similar users to gain prompt for LLaMA2
finally sh ./llama2-SFT/generate.sh to gain answers from LLM.