Awesome

GLoRE

A benchmark for evaluating the logical reasoning of LLMs

For more information, please refer to our Arxiv preprint

Datasets included:

LogiQA
ReClor
FOLIO
ConTRoL
AR-LSAT
FRACAS
HELP (The size of the HELP dataset surpasses the 25M limit, so we put it in Google Drive here. We also put a 1000 intances version of HELP in the repository.)
ProofWriter
RuleTaker
TaxiNLI
NaN-NLI
RobustLR (First 1k pieces included. The 45k full version available at Google Drive)
LogicInduction
ConCeptualCombinations

We are working on incorporating more logical reasoning datasets!

Setting-Up

pip install evals

Evaluating OpenAI Models

This repository is compatible with the OpenAI Eval library. Please download the Eval package first, and put the contents in this repository data and evals into evals/evals/registry/data/<name_of_your_eval/ and evals/evals/registry/evals/, respectively.

eg. evals/evals/registry/data/logiqa/logiqa.jsonl, evals/evals/registry/evals/logiqa.yaml

export openai api key to the environment

export OPENAI_API_KEY=<your_key>

run eval

oaieval <model_name> <data_name>

eg. oaieval gpt-3.5-turbo logiqa

Evaluating Huggingface Models

python inference.py

Contribute Your Own Dataset

We welcome your datasets incorporated in GLoRe.

Please fill free to drop us issues (in the repo) or emails (address provided in the paper) to let us know.

We would recommend you to convert your dataset into GLoRe format with the scripts provided in example_scripts/, which provides three example conversion scripts for .csv, .tsv, and .json formats respectively. But we will also handle them if you meet trouble during format conversion.

How to Cite

@misc{liu2023glore,
      title={GLoRE: Evaluating Logical Reasoning of Large Language Models}, 
      author={Hanmeng liu and Zhiyang Teng and Ruoxi Ning and Jian Liu and Qiji Zhou and Yue Zhang},
      year={2023},
      eprint={2310.09107},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}