Probing for natural language explanations.
Configure model and dataset to probe in src/conf/generate_hidden_config.yaml. Default is cross-encoder/nli-deberta-v3-base on esnli (which is just SNLI). Then run
python -m src.generate_hidden
to generate an .hdf5 file with hidden states of the model in hidden_states/.