Understanding Hallucinations in LLMs: A Graph-Based Reproduction Study (COMP4222 Course project)

Overview

We performed experiments on the pipeline of the paper "Leveraging Graph Structures to Detect Hallucinations in Large Language Models" regarding similarity threshold, embedding model, LLM, and dataset.

Group Members & Task Distribution:

• YIP Sau Lai: similarity threshold & embedding model

  • Replaced BERT with DeBERTa: pipeline/graph/make_graph_DeBERTa.py
  • For each embedding model, used thresholds: 0.60, 0.65, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95
  • Result: result/threshold&embed.tar.gz

• LAM Sum Ying: LLM for generating hallucination data

  • replaced the original LLM (LLama2) with Qwen2.5-14B to generate data with the original dataset: data_preprocess/document_generation_qwen.py
  • result: result/qwen_result

• Peng Muzi: dataset for generating hallucination data

  • replaced the original dataset with a new dataset, SciQ with the original LLM: data_preprocess/reformat_SciQ.py
  • result: result/SciQ_result

Reproduction Guidance

• similarity threshold & embedding model

  • Create result folders with corresponding names under: result/threshold&embed/
  • Copy "data" folder from pipeline/data, create "weights" and "images" folder in each result folder.
  • Run script: jobScript/threshold&embed.s

• Qwen as LLM

  • Copy sample_data.json from pipeline/data/sampled_data.json to result/qwen_result
  • Run script: jobScript/qwen_result.s (first part)
  • Create "weights" and "images" folder under result/qwen_result
  • Run script: jobScript/qwen_result.s (second part)

• SciQ as dataset

  • Run file: data_preprocess/reformat_SciQ.py
  • Run script: jobScript/SciQ_result.s (first part)
  • Create "weights" and "images" folder under result/SciQ_result
  • Run script: jobScript/SciQ_result.s (second part)

Details of the Pipeline & Results

• Please see the file "Presentation.pdf" for details
• 

References

[1] N. Nonkes, S. Agaronian, E. Kanoulas, and R. Petcu, "Leveraging Graph Structures to Detect Hallucinations in Large Language Models," in Proceedings of TextGraphs-17: Graph-based Methods for Natural Language Processing, Bangkok, Thailand, Aug. 2024, pp. https://aclanthology.org/2024.textgraphs-1.7

[2] Johannes Welbl, Nelson F. Liu, Matt Gardner, SciQ: "Crowdsourcing Multiple Choice Science Questions", Proceedings of the Workshop on Noisy User-generated Text (W-NUT) 2017. https://arxiv.org/abs/1707.06209