conda create --name gpt python=3.9 -y
conda activate gpt
# https://pytorch.org/get-started/locally/
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia -y
python -c "import torch; print(torch.__version__)"
python -c "import torch; print(torch.version.cuda)"
conda install pyg -c pyg
pip install pandas
pip install ogb
pip install transformers
pip install peft
pip install wandb
pip install sentencepiece
Download the raw texts of titles and abstracts here, unzip and move it to dataset/ogbn_arxiv_orig/.
# preprocess
python -m src.dataset.preprocess.arxiv
# check
python -m src.dataset.arxiv
Replace path to the llm checkpoints in the src/model/__init__.py, then run
python train.py --dataset arxiv --model_name graph_llm --llm_model_name 7b --gnn_model_name gat --seed 0
You may refer to run.sh for detailed commands for reproductions. The main branch is for node classification experiments and the link_prediction branch is for link prediction experiments.