FACT-AI

This is the repository for the project course FACT-AI. This contains a reproducibility study of the paper On Measuring Faithfulness or Self-consistency of Natural Language Explanations by Parcalabescu & Frank, 2024. Additionally, we added new experiments and extensions.

Installation

There are two conda environments used in this project, fact and fact2.

Install the fact environment

To install the conda environment you can run:

conda env create -f environment_fact.yml
conda activate fact
python3 -m spacy download en_core_web_sm

Finally, Wordnet from NLTK needs to be installed. This can be done by running the following python code:

import nltk
nltk.download("wordnet")

Install the fact2 environment

To install the fact2 conda environment you can run:

conda env create -f environment_fact2.yml
conda activate fact2
python3 -m spacy download en_core_web_sm

If the NLTK wordnet has not been downloaded yet (explained in the fact environment install) this needs to be done to run the code. Wordnet can be installed the following code:

import nltk
nltk.download("wordnet")

Usage

Running a test works in two steps.

1. In the file faithfulness.py, select which tests you want to use (by commenting or uncommenting a certain test in the TESTS list, line 103).
2. run python3 faithfulness.py <dataset> <model_name> <n_samples> <explainer> <explainer_max_evaluation> <loo_similarity_threshold>

Additional arguments can be

• --result_dir=<own_directory> to store the raw output in a certain directory
• --classify_pred to use the original way of using a seperate classification step to predict the label of the prompt instead of using the output of the explainer. This is only relevant for the tests that use CC-SHAP post-hoc

Not all models can be run with the same conda environment.
fact is used for: LLama2, Falcon and Mistral
fact2 is used for newer models: Falcon 3, Phi 3 and Phi 4

Running experiments

The reproducibility experiments can be done by setting the following tests: ['atanasova_counterfactual', 'atanasova_input_from_expl', 'cc_shap-posthoc', 'turpin', 'cc_shap-cot'] and running the following command.

conda activate fact
python3 faithfulness.py <dataset> <model_name> 100 "partition" 500 0.68 --result_dir="reproducibility_results"

• <model_name> is either llama2-7b-chat, falcon-7b-chat or mistral-7b-chat \
• <dataset> is either comve, esnli or disambiguation_qa

The LOO measure experiment can be done by setting the following tests: ['loo-posthoc', 'loo-cot'] and running the following command:

python3 faithfulness.py <dataset> <model_name> 100 "partition" 500 0.68 --result_dir="loo_test"

• <model_name> is either llama2-7b-chat, falcon-7b-chat or mistral-7b-chat \
• <dataset> is either comve, esnli or disambiguation_qa

The speed/time experiment for CC-SHAP and LOO is done by setting the following tests: ['cc_shap-posthoc', 'cc_shap-cot', 'loo-posthoc', 'loo-cot', 'loo-posthoc-slow', 'loo-cot-slow']
and running the command:

conda activate fact
python3 faithfulness.py comve falcon-7b-chat 100 "partition" 500 0.68 --result_dir="timing"

The alternative prompt experiment is done by setting the following tests: [cc_shap-cot-other-input, loo-cot-other-input]

python3 faithfulness.py <dataset> <model_name> 100 "partition" 500 0.68 --result_dir="alt_prompt"

-<model_name> is either falcon-7b-chat or falcon3-7B-chat

• <dataset> is either comve, esnli or disambiguation_qa

Results

There are two notebooks.

• LOO_analysis.ipynb This notebook provides the evaluation of the Leave One Out (LOO) cosine similarity measure and the determination of the similarity threshold. This notebook can be executed using the fact2 environment.
• result_analysis.ipynb In this notebook, the results for all the experiments are transformed from there raw output data (located in the results folder) into human readable dataframes. This notebook can be executed using the fact environment.