README.md

metadata

task_categories:
  - question-answering
language:
  - en
  - zh
tags:
  - biology
  - medical
size_categories:
  - 1K<n<10K
viewer: true
configs:
  - config_name: default
    data_files:
      - split: test
        path:
          - AnesBench.json
license: c-uda

The AnesBench Datasets Collection comprises three distinct datasets: AnesBench, an anesthesiology reasoning benchmark; AnesQA, an SFT dataset; and AnesCorpus, a continual pre-training dataset. This repository pertains to AnesBench. For AnesQA and AnesCorpus, please refer to their respective links: https://huggingface.co/datasets/MiliLab/AnesQA and https://huggingface.co/datasets/MiliLab/AnesCorpus.

Dataset Description

AnesBench is designed to assess anesthesiology-related reasoning capabilities of Large Language Models (LLMs). It contains 4,427 anesthesiology questions in English. Each question is labeled with a three-level categorization of cognitive demands and includes Chinese-English translations, enabling evaluation of LLMs’ knowledge, application, and clinical reasoning abilities across diverse linguistic contexts.

JSON Sample

    {
        "id": "1bb76e22-6dbf-5b17-bbdf-0e6cde9f9440",
        "choice_num": 4,
        "answer": "A",
        "level": 1,
        "en_question": "english question",
        "en_X": "option X",
        "zh_question": "中文问题",
        "zh_X": "选项X",
    }

Field Explanations

Field	Type	Description
id	string	A randomly generated ID using UUID
choice_num	int	The number of options in this question
answer	string	The correct answer to this question
level	int	The cognitive demand level of the question (1, 2, and 3 represent system1, system1.x, and system2 respectively)
en_question	string	English description of the question stem
cn_question	string	Chinese description of the question stem
en_X	string	English description of the option (X takes values from A until the total number of options is reached)
cn_X	string	Chinese description of the option (X takes values from A until the total number of options is reached)

Recommended Usage

• Question Answering: QA in a zero-shot or few-shot setting, where the question is fed into a QA system. Accuracy should be used as the evaluation metric.