RAG Hallucination Detecting By LRP

Introduction

Using LRP-based methods to detect hallucination in RAG, this is code repository for LRP4RAG: Detecting Hallucinations in Retrieval-Augmented Generation via Layer-wise Relevance Propagation.

What's LRP?

Layer-wise Relevance Propagation (LRP) is a technique that brings such explainability and scales to potentially highly complex deep neural networks.

Our Approach

Structure

baseline: code to run baselines, including 3 baselines(SelfCheckGPT, Prompt LlaMA/GPT, Finetune)
baseline_output: intermediate results after running baselines
core: code for our approach, plus data analytics and visualization
data: raw data from RAG-Truth
lrp_result_llama_7b: LRP results for llama-2-7b-chat
lrp_result_llama_13b: LRP results for llama-2-13b-chat
pdf: visualization pdf

download baseline_output.zip lrp_result_llama_7b.zip lrp_result_llama_13b.zip from NJUBox

Usage

1. run python core/llama_lrp.py to get LRP results, which will be saved in lrp_result_llama_7b
2. run python core/classifier.py --datasets ../lrp_result_llama_7b --classifier SVM to get classification result of our approach

Baselines

• SelfCheckGPT:

  • run python baseline/self_checkgpt.py
  • run python baseline/self_check_result.py

• Prompt LlaMA/GPT:

  • run python baseline/prompt_llama.py / run python baseline/prompt_gpt.py
  • run python baseline/prompt_llm_result.py

• Finetune:

  • run python baseline/llm_finetune.py

Cite

@misc{hu2024lrp4ragdetectinghallucinationsretrievalaugmented,
      title={LRP4RAG: Detecting Hallucinations in Retrieval-Augmented Generation via Layer-wise Relevance Propagation}, 
      author={Haichuan Hu and Yuhan Sun and Quanjun Zhang},
      year={2024},
      eprint={2408.15533},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2408.15533}, 
}