Dataset Card for NoiserBench

This dataset card describes NoiserBench, a comprehensive evaluation framework for analyzing the role of noise in Retrieval-Augmented Generation (RAG) systems with Large Language Models.

Dataset Details

Dataset Description

NoiserBench is a comprehensive benchmark designed to evaluate how different types of noise affect Large Language Models in Retrieval-Augmented Generation scenarios. The benchmark encompasses multiple datasets and reasoning tasks, specifically designed to analyze seven distinct noise types from a linguistic perspective. This framework reveals that noise can be categorized into two practical groups: beneficial noise (which may enhance model capabilities) and harmful noise (which generally impairs performance).

• Language(s) (NLP): English
• License: MIT
• Paper: Pandora's Box or Aladdin's Lamp: A Comprehensive Analysis Revealing the Role of RAG Noise in Large Language Models

Dataset Sources

• Repository: https://github.com/jinyangwu/NoiserBench
• Paper: https://arxiv.org/abs/2408.13533

Uses

NoiserBench is designed for:

• Evaluating the robustness of RAG systems under different noise conditions
• Analyzing how various noise types affect LLM performance in retrieval scenarios
• Benchmarking different LLM architectures and scales on noisy retrieval tasks
• Research into developing more robust and adaptable RAG solutions
• Understanding the distinction between beneficial and harmful noise in RAG contexts

Dataset Structure

The benchmark encompasses multiple datasets and reasoning tasks designed to evaluate seven distinct noise types from a linguistic perspective. The framework categorizes noise into:

1. Beneficial Noise: Types of noise that may enhance model capabilities and overall performance
2. Harmful Noise: Types of noise that generally impair LLM performance

The evaluation framework includes various reasoning tasks to comprehensively assess how different LLM architectures respond to these noise categories.

Citation

BibTeX:

@article{wu2024pandora,
  title={Pandora's Box or Aladdin's Lamp: A Comprehensive Analysis Revealing the Role of RAG Noise in Large Language Models},
  author={Wu, Jinyang and Che, Feihu and Zhang, Chuyuan and Tao, Jianhua and Zhang, Shuai and Shao, Pengpeng},
  journal={arXiv preprint arXiv:2408.13533},
  year={2024}
}

APA:

Wu, J., Che, F., Zhang, C., Tao, J., Zhang, S., & Shao, P. (2024). Pandora's Box or Aladdin's Lamp: A Comprehensive Analysis Revealing the Role of RAG Noise in Large Language Models. arXiv preprint arXiv:2408.13533.

Glossary

• RAG (Retrieval-Augmented Generation): A method that combines information retrieval with text generation to reduce hallucinations in large language models
• Beneficial Noise: Types of noise that may enhance certain aspects of model capabilities and overall performance
• Harmful Noise: Types of noise that generally impair LLM performance in RAG scenarios
• NoiserBench: The comprehensive evaluation framework established in this work

Dataset Card Contact

For questions about this dataset card or the underlying benchmark, please refer to the code repository or contact me at wu-jy23@mails.tsinghua.edu.cn.