![](https://github.com/zhuohaoyu/FreeEval/raw/main/assets/images/freeeval-logo.png)
FreeEval: A Modular Framework for Trustworthy and Efficient Evaluation of Large Language Models
Overview • Quick Start • Docs • Paper • Citation
FreeEval is a modular and extensible framework for conducting trustworthy and efficient automatic evaluations of large language models (LLMs). The toolkit unifies various evaluation approaches, including dataset-based evaluators, reference-based metrics, and LLM-based evaluators, within a transparent and reproducible framework. FreeEval incorporates meta-evaluation techniques such as human evaluation and data contamination detection to enhance the reliability of evaluation results. The framework is built on a high-performance infrastructure that enables efficient large-scale evaluations across multi-node, multi-GPU clusters, supporting both open-source and proprietary LLMs. With its focus on modularity, trustworthiness, and efficiency, FreeEval aims to provide researchers with a standardized and comprehensive platform for gaining deeper insights into the capabilities and limitations of LLMs.
To get started, first clone the repository and setup the enviroment:
git clone https://github.com/WisdomShell/FreeEval.git
cd FreeEval
pip install -r requirements.txt
All our evaluation pipelines are configured with JSON configs, including all the details and hyper-parameters. For an example, you could run ARC-Challenge with LLaMA-2 7B Chat with:
python run.py -c ./config/examples/arcc.json
For more detailed usage, please refer to our docs.
✨ If you find our work helpful, please consider citing with:
@article{yu2024freeeval,
title={FreeEval: A Modular Framework for Trustworthy and Efficient Evaluation of Large Language Models},
author={Yu, Zhuohao and Gao, Chang and Yao, Wenjin and Wang, Yidong and Zeng, Zhengran and Ye, Wei and Wang, Jindong and Zhang, Yue and Zhang, Shikun},
journal={arXiv preprint arXiv:2404.06003},
year={2024}
}