README.md

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

# AnesBench

[**Paper**](https://huggingface.co/papers/2504.02404) | [**GitHub**](https://github.com/mililab/anesbench)

# Dataset Description

**AnesBench** is designed to assess anesthesiology-related reasoning capabilities of Large Language Models (LLMs). It provides bilingual (English and Chinese) anesthesiology questions across two separate files. Each question is labeled with a three-level categorization of cognitive demands based on dual-process theory (System 1, System 1.x, and System 2), enabling evaluation of LLMs' knowledge, application, and clinical reasoning abilities across diverse linguistic contexts.

| Subset | File | Total | System 1 | System 1.x | System 2 |
|--------|------|-------|----------|-------------|----------|
| English | `anesbench_en.json` | 4,343 | 2,960 | 1,028 | 355 |
| Chinese | `anesbench_zh.json` | 3,529 | 2,784 | 534 | 211 |

## JSON Sample

**English** (`anesbench_en.json`):

```json
{
    "id": "91b5e145-57f2-5307-99e4-eafd75643de4",
    "question": "The concentration of a specific gas in solution depends on which of the following?",
    "A": "Temperature of the solution",
    "B": "Volume of the system",
    "C": "Solubility of the specific gas in that solution",
    "D": "Molecular weight of the gas",
    "choice_num": 4,
    "target": "C",
    "category": 1
}
```

**Chinese** (`anesbench_zh.json`):

```json
{
    "A": "替代治疗",
    "B": "手术治疗",
    "C": "对症治疗",
    "D": "静脉输注糖皮质激素",
    "E": "补充盐皮质激素",
    "id": "78587bd9-f3f6-4118-b6eb-95ed7c91a0ec",
    "question": "Addison病抢救的主要措施是",
    "choice_num": 5,
    "target": "D",
    "category": 1
}
```

## Field Explanations

| Field         | Type   | Description                                                                 |
|------------------|----------|-----------------------------------------------------------------------------|
| `id`             | string   | A randomly generated ID using UUID                                          |
| `question`       | string   | The question stem                                                           |
| `A`–`I`          | string   | Answer options (from `A` up to the total number of options)                 |
| `choice_num`     | int      | The number of options in this question                                      |
| `target`         | string   | The correct answer to this question                                         |
| `category`       | int      | The cognitive demand category of the question (`1` = System 1, `2` = System 1.x, `3` = System 2) |

### Cognitive Demand Categories

| Category | Label | Description |
|----------|-------|-------------|
| 1 | **System 1** | Fast, intuitive recall of factual knowledge |
| 2 | **System 1.x** | Pattern recognition and application of learned rules |
| 3 | **System 2** | Deliberate, analytical clinical reasoning |

## 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.

## Usage

To evaluate a model on AnesBench, you can use the evaluation code provided in the [official repository](https://github.com/MiliLab/AnesSuite).


## Citation

If you find AnesBench helpful, please consider citing the following paper:

```latex
@inproceedings{
  feng2026anessuite,
  title={AnesSuite: A Comprehensive Benchmark and Dataset Suite for Anesthesiology Reasoning in {LLM}s},
  author={Xiang Feng and Wentao Jiang and Zengmao Wang and Yong Luo and Pingbo Xu and Baosheng Yu and Hua Jin and Jing Zhang},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=iKRQMeC7yO}
}
```