Graph Reasoning Prefix for Language Model Question Answering

Dependencies

pip install numpy tqdm
pip install torch torchvision
pip install transformers==4.11 nltk spacy
pip install wandb
pip install tensorboardx
pip install tensorboard

# for torch-geometric
pip install torch-scatter
pip install torch-cluster
pip install torch-sparse
pip install torch-spline-conv
pip install torch-geometric

Download data

You can download all the preprocessed data here. Download data_preprocessed.zip and data_preprocessed_biomed.zip into the top-level directory of this repo and unzip them. Move the medqa_usmle and ddb folders into the data/ directory.

The resulting file structure should look like this:

├── README.md
├── data/
    ├── cpnet/                 (prerocessed ConceptNet)
    ├── csqa/
        ├── train_rand_split.jsonl
        ├── dev_rand_split.jsonl
        ├── test_rand_split_no_answers.jsonl
        ├── statement/             (converted statements)
        ├── grounded/              (grounded entities)
        ├── graphs/                (extracted subgraphs)
        ├── ...
    ├── obqa/
    ├── medqa_usmle/
    └── ddb/

Training

CUDA_VISIBLE_DEVICES=0 ./run_prefix_fusion.sh csqa --data_dir data/ --inhouse False

Evaluating

CUDA_VISIBLE_DEVICES=0 ./eval_prefix_fusion.sh csqa --data_dir data/ --inhouse False --load_model_path /path/to/checkpoint