# Halo: Estimation and Reduction of Hallucinations in Open-Source Weak Large Language Models

**Authors:** Mohamed Elaraby, Mengyin Lu, Jacob Dunn, Xueying Zhang, Yu Wang, Shizhu Liu, Pingchuan Tian, Yuping Wang, Yuxuan Wang  
**arXiv:** [2308.11764](https://arxiv.org/abs/2308.11764)  
**Date:** 2023-08-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.11764v4/

## Key points

- Authors enhance performance of Large Language Models (LLMs) through knowledge injection
- Utilize entity triplets and summaries from Wikipedia API to create 54K training samples for NBA domain
- Preserve triplet format for naturalness in generated responses
- Model trained using special token "TRUE_FACT:" and causal language model objective
- Experiment with two settings for knowledge injection: Intermediate tuning and Combined tuning
- Evaluate effectiveness of techniques and knowledge retention during intermediate finetuning stages
- Introduce HaloCheck, a lightweight BlackBox framework for quantifying hallucinations in LLMs
- Compare with selfcheckGPT-NLI to show efficiency in detecting contradictions in responses
- Contribute insights into reducing hallucinations in low-parameter LLMs and introduce novel evaluation framework
- Pave way for future research to expand approaches across domains and improve model performance

## Summary

The authors of this study build upon previous research to enhance the performance of Large Language Models (LLMs) through knowledge injection. By utilizing entity triplets and summaries extracted from Wikipedia API, they create a set of 54K training samples for the NBA domain. Unlike previous approaches, they preserve the triplet format to maintain naturalness in generated responses. The model is trained using a special token "TRUE_FACT:" and a causal language model objective due to its decoder-only architecture. Two settings for knowledge injection are experimented with: Intermediate tuning where finetuning is done exclusively on knowledge text before SFT data, and Combined tuning where both types of data are jointly finetuned. The effectiveness of these techniques is evaluated along with knowledge retention during intermediate finetuning stages. Additionally, the authors introduce HaloCheck, a lightweight BlackBox framework for quantifying hallucinations in LLMs without requiring extensive computational resources or question generation modules. Comparisons with selfcheckGPT-NLI show its efficiency in detecting subtle contradictions within sampled responses through quantitative and qualitative analyses. This study contributes valuable insights into reducing hallucinations in low-parameter LLMs and introduces a novel framework for evaluating hallucination severity in generated responses. It also paves the way for future research to expand these approaches across multiple domains and improve model performance in challenging tasks.