InterrogateLLM - Hallucination Detection

This repository contains the code implementation of the InterrogateLLM method as described in the paper "InterrogateLLM: Zero-Resource Hallucination Detection in LLM-Generated Answers". The InterrogateLLM method is designed to detect hallucinations in large language models.

• Project page for the paper "InterrogateLLM: Zero-Resource Hallucination Detection in LLM-Generated Answers"

Requirements

• Python 3.x
• Required Python packages can be installed via pip install -r requirements.txt

Datasets

To run the experiments, you need to download the following datasets:

• The Movies Dataset
• Books Dataset
• Global Country Information (GCI) Dataset 2023

Preprocess data:

run:

• Movies:

  python datasets/imdb_dataset.py

• Books:

  python datasets/books_dataset.py

  Usage

  To run the experiments, use the following command:

python run_experiments.py --dataset_name=<movies/books/world> --ans_model=<gpt/llamaV2-7/llamaV2-13> --embedding_model_name=<ada002/sbert>

• --dataset_name: Specify the dataset on which the experiments will be run (movies, books, or world).
• --ans_model: Specify the language model to use for answering queries (gpt, llamaV2-7, or llamaV2-13).
• --embedding_model_name: Specify the embedding model to use for checking similarity between the reconstructed question and the original question (ada002 or sbert).

Example

To run example use the following command:

python run_example.py --ans_model=<gpt/llamaV2-7/llamaV2-13> --embedding_model_name=<ada002/sbert> --reconstruction_models=<gpt,llamaV2-7,llamaV2-13> --iterations=<number>

• --ans_model: Specify the language model to use for answering queries (gpt, llamaV2-7, or llamaV2-13).
• --embedding_model_name: Specify the embedding model to use for checking similarity between the reconstructed question and the original question (ada002 or sbert).
• --reconstruction_models: Specify the language models to employ for reconstructing the query from the predicted answer. The options include permutations of gpt, llamaV2-7, and llamaV2-13.
• --iterations: number of iterations to reconstruct the query for each model.

To examine a different query, modify the query variable along with the corresponding few-shot example in the run_example.py file.

Citation

If you use this code or method in your research, please cite the original paper:

@article{yehuda2024search,
      title={In Search of Truth: An Interrogation Approach to Hallucination Detection}, 
      author={Yakir Yehuda and Itzik Malkiel and Oren Barkan and Jonathan Weill and Royi Ronen and Noam Koenigstein},
      year={2024},
      eprint={2403.02889},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}