Interpretability using PEGs

This is the repository containing all relevant code for the project “Parse Probing in PEGs”.

Running the code

All experiments can be run through main.py. The behavior is controlled by the config.yaml file.

python main.py

The task run by main.py will be determined by the experiment.task property in the config file. The possible values for this are:

To train a model:

train_model

To train a probe:

train_probe

To run an intervention:

run_intervention (Not implemented yet)