https://github.com/ruixiangcui/AGIEval?tab=readme-ov-file

https://github.com/ruixiangcui/AGIEval?tab=readme-ov-file

Last synced: 13 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/ruixiangcui/AGIEval?tab=readme-ov-file
• Owner: ruixiangcui
• License: mit
• Created: 2023-03-26T10:37:46.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2024-06-13T14:20:51.000Z (5 months ago)
• Last Synced: 2024-08-02T06:19:35.876Z (3 months ago)
• Language: Python
• Homepage:
• Size: 6.87 MB
• Stars: 674
• Watchers: 9
• Forks: 45
• Open Issues: 7
• Metadata Files:
  • Readme: README.md
  • License: LICENSE
  • Code of conduct: CODE_OF_CONDUCT.md
  • Security: SECURITY.md

Awesome Lists containing this project

• awesome-foundation-model-leaderboards - AGIEval - centric benchmark to evaluate the general abilities of foundation models in tasks pertinent to human cognition and problem-solving. | (Model Ranking / Text)

README

        
# AGIEval

This repository contains information about AGIEval, data, code and output of baseline systems for the benchmark.

# Introduction

AGIEval is a human-centric benchmark specifically designed to evaluate the general abilities of foundation models in tasks pertinent to human cognition and problem-solving. 

This benchmark is derived from 20 official, public, and high-standard admission and qualification exams intended for general human test-takers, such as general college admission tests (e.g., Chinese College Entrance Exam (Gaokao) and American SAT), law school admission tests, math competitions, lawyer qualification tests, and national civil service exams. 

For a full description of the benchmark, please refer to our paper: [AGIEval: A Human-Centric Benchmark for

Evaluating Foundation Models](https://arxiv.org/pdf/2304.06364.pdf).

# Tasks and Data

[We have updated the dataset to version 1.1.](data/v1_1) The new version updated Chinese Gaokao (chemistry, biology, physics) datasets with questions from 2023 and addressed annotation issues. To facilitate evaluation, now all multi-choice question (MCQ) tasks have one answer only (Gaokao-Physics and JEC-QA used to have multi-label answers). AGIEval-en datasets remain the same as Verison 1.0. The new version's statistics are as follows:

AGIEval v1.1 contains 20 tasks, including 18 MCQ tasks and two cloze tasks (Gaokao-Math-Cloze and MATH). You can find the full list of tasks in the table below.

![The datasets used in AGIEVal](AGIEval_tasks.png)

You can download all post-processed data in the [data/v1_1](data/v1_1) folder. All usage of the data should follow the license of the original datasets. 

The data format for all datasets is as follows:

```

{

    "passage": null,

    "question": "设集合 $A=\\{x \\mid x \\geq 1\\}, B=\\{x \\mid-1-1\\}$", 

        "(B)$\\{x \\mid x \\geq 1\\}$", 

        "(C)$\\{x \\mid-1