This is the source code for all the experiments pertaining to the paper "Uncertainty Quantification in Natural Language Explanations of Language Models" accepted at AISTATS 2024.

As mentioned in the paper, We conduct experiments using three math word problem and two commonsense reasoning benchmark datasets. i) the GSM8K dataset that comprises several math word problems, ii) the SVAMP dataset contains math word problems with varying structures, iii) the ASDiv dataset consisting of diverse math word problems, iv) the StrategyQA requires a language model to deduce a multi-step reasoning strategy to answer questions and v) the Sports Understanding dataset, which is a specialized evaluation set from the BIG-bench that involves determining whether a sentence relating to sports is plausible or implausible.

The processed files are available in data directory.

The notebook 'EDA.ipynb' loads datasets and does pre-processing of questions. Moreover, also prepends the handcrafted prompts to the question to generate natural language explanations. The final output parquet files are written to data directory. The columns are as follows.

experiment_1_question - Verbalized Chain Of Thought question experiment_2_question - Feature Importance Explanation question experiment_3_question - Verbalized Feature Importance question experiment_4_question - Paraphrased questions for Sample Probing CoT Explanations experiment_5_question - Paraphrased questions for Sample Probing Feature Importance Explanations experiment_8_question - Model Probing CoT Explanation question experiment_9_question - Model Probing Feature Importance Explanation question

The script uncertainty-all.py takes a dataset file path as argument, and gets responses from GPT models using OpenAI's API's. Make sure you update the api key and model name in the script accordingly.

Example

python uncertainty-all.py data/gsm8k_100/input.parquet

For math word problem datasets, run the notebook "Math Word Problem Analysis.ipynb" to compute verbalized, sample probing and model probing uncertainty scores of explanations.

For commonsense reasoning datasets, run the notebook "Commonsense Reasoning Analysis.ipynb" to compute verbalized, sample probing and model probing uncertainty scores of explanations.

For math word problem datasets, run the notebook "Math Word Problem Analysis.ipynb" to compute faithfulness of token importance explanations, and CoT dataset faithfulness.

For commonsense reasoning datasets, run the notebook "Commonsense Reasoning Analysis.ipynb" to compute faithfulness of token importance explanations, and CoT dataset faithfulness.

The notebook Plots.ipynb has plotting code for the results shown in the paper.

The final processed files for all 5 datasets, and 2 models, can be accessed here.

import pandas as pd

data_all = {
    "GSM8K": {
        "text-davinci-003": pd.read_parquet("data/gsm8k_100/faithfulness-scores-text-davinci-003.parquet"),
        "gpt-3.5-turbo": pd.read_parquet("data/gsm8k_100/faithfulness-scores-gpt-3.5-turbo.parquet"),
    },
    "SVAMP": {
        "text-davinci-003": pd.read_parquet("data/svamp_100/faithfulness-scores-text-davinci-003.parquet"),
        "gpt-3.5-turbo": pd.read_parquet("data/svamp_100/faithfulness-scores-gpt-3.5-turbo.parquet"),
    },
    "ASDiv": {
        "text-davinci-003": pd.read_parquet("data/asdiv_100/faithfulness-scores-text-davinci-003.parquet"),
        "gpt-3.5-turbo": pd.read_parquet("data/asdiv_100/faithfulness-scores-gpt-3.5-turbo.parquet"),
    },
    "StrategyQA": {
        "text-davinci-003": pd.read_parquet("data/strategyqa_100/faithfulness-scores-text-davinci-003.parquet"),
        "gpt-3.5-turbo": pd.read_parquet("data/strategyqa_100/faithfulness-scores-gpt-3.5-turbo.parquet"),
    },
    "Sports Understanding": {
        "text-davinci-003": pd.read_parquet("data/sportsunderstanding_100/faithfulness-scores-text-davinci-003.parquet"),
        "gpt-3.5-turbo": pd.read_parquet("data/sportsunderstanding_100/faithfulness-scores-gpt-3.5-turbo.parquet"),
    },
}