RefuteBench: Evaluating Refuting Instruction-Following for Large Language Models

Introduction

RefuteBench evaluates how LLMs react to refuting instructions, demonstrating how stubborn are LLMs against user feedback. For more details, please check our paper.

Dependency

Tutorial

Bash scripts for reproduce our results can be found in "scripts/". Our implementation is mostly based on OPENAI api and VLLM. For open-source models like LLaMA-2 family, Mistral and Alpaca, we use VLLM to first serve the model and then send queries to the server. For close-source models like ChatGPT, GPT4 and Claude-2, we directly send queries to Offical API endpoints.

💡 Note

Data and code will be uploaded within the next few days. Stay tuned! [Update 2024.3.7] We upload code and scripts to reproduce our benchmark.

[Update 2024.2.29] The data is uploaded in data/*. Our benchmark contains three tasks, namely machine translation, question answering and email writing. Then, we have two scenarios: single-feedback and multi-feedback. The single-feedback scenario is a clean setting to study problems like memory loss and instruction following, and the multi-feedback setting mimics the real-world application. We recommend to using the multi-feedback directly to test your LLMs.

☕️ Citation

If you find this repository helpful, please consider citing our paper:

@misc{yan2024refutebench,
      title={RefuteBench: Evaluating Refuting Instruction-Following for Large Language Models}, 
      author={Jianhao Yan and Yun Luo and Yue Zhang},
      year={2024},
      eprint={2402.13463},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}