This repo tests an online version of LEACE to erase spurious heuristic concepts from BERT representations during finetuning on the MNLI dataset. All the models are evaluated on the HANS dataset, which evaluates the robustness of a model with respect to these spurious heuristics.
Create a Python >= 3.10 environment and install the requirements:
pip install -e .One can finetune a BERT model on the MNLI dataset and evaluate the resulting model on the HANS dataset with the command:
python cmd/train.py --concept-erasure ERASURE_METHOD --include_sublayerswhere --concept_erasure and --include_sublayers are optional flags controlling the concept erasure technique used, ERASURE_METHOD` is an optional argument assuming one of the following values:
| ERASURE_METHOD | Description |
|---|---|
leace-cls |
LEACE applied to the representation of each CLS token. |
leace-flatten |
LEACE with orth flag applied to sequence representation. |
This model can be evaluated (with or without concept erasure) via the command:
python cmd/evaluate_model.py XYZ --concept-erasure ERASURE_METHOD --include_sublayerswhere XYZ is the ID of the model to be evaluated (assign in the ./results directory) , where --concept_erasure and --include_sublayers are optional flags and ERASURE_METHOD is an optional argument assuming one of the above values.
Development dependencies can be installed with:
pip install -e .[dev]