You Believe Your LLM is Not Delusional? Think Again!

This repository contains the code accompanying our paper titled "You Believe Your LLM is Not Delusional? Think Again! A Study of LLM Hallucination on Foundation Models under Perturbation," which is currently under review.

Overview

In this work, we present an evaluation framework designed to detect hallucinations in Large Language Models (LLMs). Our approach involves:

• Query Perturbation: Applying controlled perturbations at different levels (character, word, and sentence) to the input queries.
• Consistency Score Calculation: Measuring the consistency between the responses generated for the original query and the perturbed query to identify potential hallucination scenarios.

Repository Structure

This repository is organized into the following sections:

1. Prepare Data

• Description: Contains scripts to download datasets from Hugging Face Hub and apply various perturbations to the queries at character, word, and sentence levels.
• Files and Modules:
  • perturbations: Module with functions to introduce character-level, word-level, and sentence-level perturbations.
  • main.py: Script to download data from Hugging Face Hub and apply perturbations.

2. Generate Responses

• Description: Includes code for querying LLMs and obtaining responses for both original and perturbed queries.
• Files and Modules:
  • generate_response.py: Function to format LLM call.
  • main.py: Script to generate responses using LLM.

3. Generate Scores

• Description: Provides tools for calculating hallucination scores by comparing responses to original and perturbed queries.
• Files and Modules:
  • utils.py: Contains the model to return scores.
  • main.py: Script to generate scores.

Usage Instructions

1. Prepare the environment: Ensure you have the required dependencies installed by running:

   pip install -r requirements.txt

2. Run the data preparation pipeline:

   python -m prepare_data.main

3. Generate LLM responses:

   python -m generate_responses.main

4. Calculate hallucination scores:

   python -m generate_scores.main

Contact

For any questions or discussions, please feel free to reach out to the authors:

Name	Email Address
Anirban Saha	[email protected]
Binay Gupta	[email protected]
Anirban Chatterjee	[email protected]
Kunal Banerjee	[email protected]