🌟 HalluShift

This repository contains the source code for the IJCNN 2025 paper HalluShift: Measuring Distribution Shifts towards Hallucination Detection in LLMs by Sharanya Dasgupta, Sujoy Nath, Arkaprabha Basu, Pourya Shamsolmoali, and Swagatam Das.

---

🛠️ Model Preparation

1. 📦 Setup Environment
   • Install Python 3.10.12 and the necessary packages from requirements.txt.
   • For easily managing different python versions, we recommend using conda.
   • Create a new environment in conda and install necessary python packages:

     conda create -n hallushift python=3.10.12 -y
     conda activate hallushift
     pip install -r requirements.txt

2. 🔑 Create HF Access Token
   • Login to huggingface or create an account if you don't have already.
   • From the settings create a new access token with WRITE access.
   • Open the hal_detection.py and paste your access token at line 15, hf_token = "<INPUT_YOUR_HF_ACCESS_TOKEN>"

---

📊 Generating Responses and Hallucination Detection

1. 📁 Setup Necessary Directoriess
   Create a folder to save:

   • LLM-generated answers and Ground truth labels for model-generated content
   • Features for training classifiers

   mkdir results

   Create another folder, where you need to place BleuRT models in the next step. And llm models will be automatically saved while running the script.

   mkdir models
   

2. ✅ Ground Truth Evaluation
   Since generated answers lack explicit ground truth, we use BleuRT to evaluate truthfulness.

   • To install BleuRT run:

   pip install --upgrade pip  # ensures that pip is current
   git clone https://github.com/google-research/bleurt.git
   cd bleurt
   pip install .
   

   • We are using 12-layer distilled model for faster inference, which is ~3.5X smaller.
   • Download the model and save it in the ./models folder:

   wget https://storage.googleapis.com/bleurt-oss-21/BLEURT-20-D12.zip
   unzip BLEURT-20-D12.zip
   mv BLEURT-20-D12 models/. # Move the bleurt model folder to models directory
   

   • If you want to use any different model please refer to BleuRT repository.

3. 🔍 Hallucination Detection for TruthfulQA
   To perform hallucination detection on the TruthfulQA dataset run the following command:

   python hal_detection.py --dataset_name truthfulqa --model_name llama2_7B 

   • dataset_name: Choose from truthfulqa, triviaqa, tydiqa, coqa, haluevalqa, haluevaldia, haluevalsum.
   • model_name: Choose from llama2_7B, llama3_8B, opt6.7B, vicuna_7B or Qwen2.5_7B.
   • To use a different LLM, add its corresponding path to the MODELS_NAMES dictionary in hal_detection.py.
   • To use a different dataset, update the load_dataset_by_name function in hal_detection.py. Ensure that it returns a datasets.Dataset object.

   Note: If you encounter memory errors, consider reducing the number of workers using the --num_workers parameter. For example:

   python hal_detection.py --dataset_name truthfulqa --model_name llama2_7B --num_workers 1

   Refer to Section IV of the paper for implementation details.

---

• 🎮 Demo To quickly evaluate the pre-trained model on the TruthfulQA dataset (which has already been processed and inferred using the LLaMA-2 7B model), run the following command:

  cd demo
  python demo.py

  The pre-trained model and the processed dataset are provided in the demo folder for easy evaluation.