Skip to content

Latest commit

 

History

History
52 lines (36 loc) · 3.98 KB

README.md

File metadata and controls

52 lines (36 loc) · 3.98 KB
Raw

PRE: Peer Review Evaluator

This is a general framework used on evaluating the performance of large language models (LLMs) based on the peer review mechanism among LLMs.

Get Started

Install nessary package:

pip install -r requirements.txt

Run toy example:

python main.py --config config/main.yaml

Framework

Following figure shows the framework of our evaluation framework.

overview framework

Our whole framework could be divided into three main modules

  • Qualification exam module: Conduct qualification exam, to filter out in-qualified LLMs. You could modify exam.yaml to config the setting of this module. Some key config parameters:
    • source: same / others. same denotes the exam data comes from the samples of LLMs' task responses, while others denotes the exam data is collected from other tasks.
    • mode: pointwise / pairwise. Determine the exam task's format.
    • conduct_reference_exam: whether or not to conduct reference exam, i.e., compare the evaluator candidate LLM's responses with the ground truth (need to be provided by user)
    • conduct_inner_consistency_exam: whether or not to conduct inner-consistency exam, i.e., compare the candidate LLM's responses on the given prompt template with another new prompt template (need to be provided by user)
    • metric_pointwise: EM (exact match) / MSE (mean square error), the evaluation metric of the LLMs' exam performance on pointwise mode
  • Peer review module: Collect the responses of all evaluatee LLMs to the given tasks. Then, each qualified reviewer LLM is required to rate the outputs in the response pool set. You could modify evaluatees.yaml and evaluators.yaml to config the setting of the evaluatee LLMs and evaluator LLMs, also you could modify data.yaml to config the setting of task data loader. It is worth noting that our LLM APIs support user-customed. You could inherit the abstract class LLM_API to implement a custom interface. You just need to register this interface class in the API_type2class_list before using it.
  • “Chair” decision module: Aggregate the ratings of evaluator LLMs and generate the final evaluation results. You could modify eval.yaml to config the setting of this module. Some key config parameters:
    • strategy: full / ELO / Glicko. full denote using the whole pointwise / pairwise set of evaluation tasks to generate the final results. While ELO and Gilcko are widely used in chess gaming, they only need to use some samples to obtain the final results.
    • mode: pointwise / pairwise. Determine the exam task's format.
    • weighted_method: uniform / exp / log / poly. The weights of each evaluator, which are computed based on the performance of LLMs in qualification exam.

There are several advantages of our framework:

  • Supports Disconnected Connections: to tackle the potential unstable situations of the LLM APIs
  • User code-free: user is able to run the project just after supplementing their own task data and configing the yaml config files, no need to modify the codes
  • Customed-friendly: user could customize their special extension if they have some expansion usage requirement. Our framework is welcome for these personalized requirments
  • Support mutli-processing: Our framework support multi-processing, user could run many APIs at the same time.

Citation

If you used our evaluation framework, please cite with the following information.

Zhumin Chu, Qingyao Ai, Yiteng Tu, Haitao Li, Yiqun Liu. "PRE: a peer review evaluator framework for automatic LLM evaluation." 2023.