Note: You should run the scripts in ./my_project_gpt2, since the Llama2 version of our model is still buggy
| Method | CSQA-train | CSQA-test |
|---|---|---|
| LMGNN-GPT2-Small | 20.9% | 21.5% |
| Vanilla GPT2-Small | 5.7% | 4.1% |
| Full-finetune GPT2-Small | 70.1% | 20.5% |
The test data and checkpoint can be accessed via Google Drive.
You should follow the steps described in install_steps.txt to install the Python environment.
You can use preprocess.py to preprocess the dataset for training. Note that the Llama2 model file and the original datasetus are not contained in this repository.
You can use my_project_gpt2/main.py to train the model.
You can use baseline_lmgnn.py and baseline_full_ft.py to evaluate our model and baseline.
Our trained checkpoints will be uploaded to Github.