⏬ Data • 📃 Paper • 🏆 Leaderboard
中文 | English

AC-EVAL presents a thorough evaluation suite for Large Language Models (LLMs) focusing on ancient Chinese, covering eras from the Pre-Qin period to the Qing dynasty. This suite includes 3245 multi-choice questions across 3 levels of difficulty and 13 diverse tasks, as shown below. Please check our paper for more details.

Our aim is to facilitate the assessment of LLMs' capabilities in understanding and processing ancient Chinese language and knowledge.

Leaderboard

Our leaderboard, which is updated regularly, showcases both zero-shot and five-shot accuracies of various models across two distinct settings: Answer-Only (AO) and Chain-of-Thought (COT). The breakdown results are shown in breakdown_results.xlsx

Zero-shot AO

Model	General Historical Knowledge	Short Text Understanding	Long Text Understanding	Average
ERNIE-Bot 4.0	77.54	68.11	66.42	70.69
GLM-4	76.63	66.66	67.70	70.33
Qwen-max	70.77	64.88	63.84	67.50
GLM-3-Turbo	75.21	60.52	59.77	65.17
Qwen-72B-Chat	71.25	61.48	59.80	64.18
Yi-34B-Chat	72.66	61.33	58.36	64.12
Qwen-14B-Chat	69.51	56.53	57.38	61.14
ERNIE-Bot	68.81	57.80	51.47	59.36
GPT-4	66.11	55.11	47.38	56.20
Qwen-7B-Chat	62.74	48.76	44.97	52.16
Yi-6B-Chat	60.70	47.79	39.49	51.33
Baichuan2-7B-Chat	64.38	46.77	40.33	50.49
Baichuan2-13B-Chat	65.57	49.24	35.40	50.07
ChatGLM3-6B	58.04	43.01	39.73	46.93
Xunzi-Qwen-Chat	60.20	44.31	30.87	45.13
GPT-3.5 Turbo	53.50	43.72	36.94	44.72
LLaMA2-70B	33.55	36.29	30.72	33.54

Five-shot AO

Model	General Historical Knowledge	Short Text Understanding	Long Text Understanding	Average
ERNIE-Bot 4.0	75.69	69.59	66.12	70.47
GLM-4	74.89	65.48	69.07	69.81
Qwen-max	75.29	65.48	66.99	69.25
GLM-3-Turbo	72.99	59.48	59.66	64.04
Qwen-72B-Chat	71.67	61.30	57.07	63.35
ERNIE-Bot	68.81	57.62	50.36	58.93
GPT-4	65.91	58.07	48.36	57.45
Qwen-14B-Chat	70.60	53.73	45.91	56.75
Yi-34B-Chat	66.62	52.57	41.90	53.70
Baichuan2-7B-Chat	63.37	45.91	39.94	49.74
Baichuan2-13B-Chat	63.75	45.86	32.74	47.45
Qwen-7B-Chat	61.42	45.98	30.78	46.06
ChatGLM3-6B	55.74	42.92	38.45	45.71
GPT-3.5 Turbo	53.99	43.21	36.40	44.54
Xunzi-Qwen-Chat	51.30	41.25	29.84	40.80
Yi-6B-Chat	55.76	35.97	28.48	40.07

Zero-shot COT

Model	General Historical Knowledge	Short Text Understanding	Long Text Understanding	Average
Qwen-max	75.10	66.72	61.03	67.62
Qwen-72B-Chat	74.79	65.25	56.78	65.61
Qwen-14B-Chat	67.51	54.64	46.12	56.09
Qwen-7B-Chat	61.54	44.97	40.21	48.91

Five-shot COT

Model	General Historical Knowledge	Short Text Understanding	Long Text Understanding	Average
Qwen-max	74.30	65.94	61.46	67.23
Qwen-72B-Chat	71.79	61.62	57.66	63.69
Qwen-14B-Chat	67.49	51.51	39.93	52.97
Qwen-7B-Chat	59.37	47.71	35.36	47.48

Data

Download

• Method 1: Download the zip file (you can also simply open the following link with the browser):

  wget https://huggingface.co/datasets/yuting-wei/aceval/resolve/main/aceval.zip
  

  then unzip it and you may load the data with pandas:

  import os
  import pandas as pd
  
  File_Dir="aceval"
  test_df=pd.read_csv(os.path.join(File_Dir,"test","art_and_cultural_heritage.csv"))

• Method 2: Directly load the dataset using Hugging Face datasets:

  from datasets import load_dataset
  
  dataset=load_dataset(r"yuting-wei/aceval",name="art_and_cultural_heritage")

Data Format

To facilitate usage, we have organized the supercategory labels and English/Chinese names corresponding to 13 subjects. Please refer to subject_mapping.json for details. The format is:

{
    "art_and_cultural_heritage": {
      "English": "Art and Cultural Heritage",
      "Chinese": "艺术和文化遗产",
      "Supercategory": "General Historical Knowledge"
    },
    ...
    "filename":{
        "English": English Name,
        "Chinese": Chinese Name,
        "Supercatagory": Supercatagory Label (General Historical Knowledge/Short Text Understanding/Long Text Understanding)"
    }
}

Each subject consists of two splits: dev and test. The dev set per subject consists of five exemplars with explanations for few-shot evaluation. The test set is for model evaluation. Labels on the test split are not released, users are required to submit their results to automatically obtain test accuracy. How to submit?

Below is a dev example from art and cultural heritage:

Question	A	B	C	D	Answer	Explanation
五代南唐时期著名画家顾闳中的绘画名作是？(The famous painting masterpiece of Gu Hongzhong, a famous painter in the Southern Tang Dynasty during the Five Dynasties, is?)	《女史箴图》(Admonitions of the Instructress to the Court Ladies)	《五牛图》(Five Buffaloes)	《簪花仕女图》(Ladies with Flowers)	《韩熙载夜宴图》(Han Xizai Giving a Night Banquet)	D	顾闳中的绘画名作是《韩熙载夜宴图》。《五牛图》是韩滉的作品，《簪花仕女图》是周昉的作品，《女史箴图》是顾恺之的作品。 (The famous painting by Gu Hongzhong is 'Han Xizai Giving a Night Banquet.' 'Five Buffaloes' is a work by Han Huang, 'Ladies with Flowers' is by Zhou Fang, and 'Admonitions of the Instructress to the Court Ladies' is by Gu Kaizhi.)

How to Evaluate on AC-EVAL

We implemented automatic answer extraction using regular expressions, the evaluation code for each model is located in the src directory.

We use the following prompt when evaluating the models:

answer-only prompt

Zero-shot AO

以下是中国古代{主题}领域的单项选择题，请直接给出正确答案对应的选项。

{测试题目}
A. {选项A}
B. {选项B}
C. {选项C}
D. {选项D}
答案：

Few-shot AO

以下是中国古代{主题}领域的单项选择题示例。在查看这些示例之后，请直接给出接下来一道题目的正确答案所对应的选项。

示例1：{题目1}
A. {选项A}
B. {选项B}
C. {选项C}
D. {选项D}
答案：A

[k-shot demo, note that k is 0 in the zero-shot case]

{测试题目}
A. {选项A}
B. {选项B}
C. {选项C}
D. {选项D}
答案：

chain-of-thought prompt

Zero-shot COT

以下是中国古代{主题}领域的单项选择题，请逐步分析并给出正确答案对应的选项。

{测试题目}
A. {选项A}
B. {选项B}
C. {选项C}
D. {选项D}
答案：

Few-shot COT

以下是中国古代{主题}领域的单项选择题示例。在查看这些示例之后，请逐步分析接下来一道题目并给出正确答案所对应的选项。

示例1：{题目1}
A. {选项A}
B. {选项B}
C. {选项C}
D. {选项D}
答案解析：
让我们逐步分析。{解析过程}
所以答案是A。

[k-shot demo, note that k is 0 in the zero-shot case]

{测试题目}
A. {选项A}
B. {选项B}
C. {选项C}
D. {选项D}
答案：

How to Submit

You need to first prepare a UTF-8 encoded JSON file with the following format, please refer to submission_example.json for details.

{
    "historical_facts": {
        "0": "A",
        "1": "B",
        "2": "B",
        ...
    },
    
    "subject_name":{
    "0":"ans_0",
    "1":"ans_1",
    ...
    }
    ....
}

Then, you can submit the prepared JSON file to the email (yuting_wei@bupt.edu.cn). Please include the type of experiment you conducted in the subject of your email, using one of the following file labels: [zero-shot-AO, few-shot-AO, zero-shot-COT, few-shot-COT].

TODO

• add evaluation code into src
• add breakdown results
• incorporate into Hugging Face datasets

Licenses

This work is licensed under a MIT License.

The AC-EVAL dataset is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Citation

Please cite our paper if you use our dataset.

@inproceedings{wei-etal-2024-ac,
    title = "{AC}-{EVAL}: Evaluating {A}ncient {C}hinese Language Understanding in Large Language Models",
    author = "Wei, Yuting  and Xu, Yuanxing  and Wei, Xinru  and Yang, Simin  and Zhu, Yangfu  and Li, Yuqing  and Liu, Di  and Wu, Bin",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    year = "2024",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.87",
    pages = "1600--1617",
}

Acknowledgements

This project was inspired by and based on the structure of C-Eval. We are grateful for this work and would like to acknowledge their significant contributions to the community.