🕵️EasyJudge: an Easy-to-use Tool for Comprehensive Response Evaluation of LLMs

• Lightweight Usage Model:EasyJudge is built to minimize dependency requirements, offering a simple installation process and precise documentation. Users can initiate the evaluation interface with only a few basic commands.

• Comprehensive Evaluation Tool: EasyJudge offers a highly customizable interface, allowing users to select evaluation scenarios and flexibly combine evaluation criteria based on their needs. The visualization interface has been carefully designed to provide users with an intuitive view of various aspects of the evaluation results.

• Efficient Inference Engine: EasyJudge employs model quantization, memory management optimization, and hardware acceleration support to enable efficient inference. As a result, EasyJudge can run seamlessly on consumer-grade GPUs and even CPUs.

System Overview

Model

EasyJudge is now available on huggingface-hub: 🤗 4real/EasyJudge_gguf

Quick Start

(Example of Deploying on AutoDL Cloud Server)

Deploy ollama

1. Start the installation software on autodl

export OLLAMA_MODELS=/root/autodl-tmp/models
curl -fsSL https://ollama.com/install.sh | sh

2. Start the service

ollama serve

3. Import EasyJudge models

Modify the path after from in each Modelfile to the local path where the model is downloaded from huggingface.

ollama create PAIRWISE -f /root/autodl-tmp/EasyJudge/Modelfile/PAIRWISE.Modelfile
ollama create POINTWISE -f /root/autodl-tmp/EasyJudge/Modelfile/POINTWISE.Modelfile

Environment Configuration

(EasyJudge uses the environment PyTorch 2.3.0, Python 3.12 (ubuntu22.04), and Cuda 12.1.)

1. Create conda environment

conda create -n EasyJudge
conda init
conda activate EasyJudge

2. Install specified Python packages in bulk

pip install -r requirements.txt

Run the Program

To start the application, use the following command to run main.py with specific server configurations:

streamlit run main.py --server.address=127.0.0.1 --server.port=6006 --server.enableXsrfProtection=false

Citation

Please cite the repo or the paper if the model/code/resource/conclusion in this repo is helpful to you.

@article{li2024easyjudge,
  title={EasyJudge: an Easy-to-use Tool for Comprehensive Response Evaluation of LLMs},
  author={Li, Yijie and Sun, Yuan},
  journal={arXiv preprint arXiv:2410.09775},
  year={2024}
}

Acknowledge❤️

We acknowledge these works for their public codes: LLaMA-Factory, llama.cpp, ollama, auto-j, JudgeLM.