GAOKAO-Bench是一个以中国高考题目为数据集,测评大模型语言理解能力、逻辑推理能力的测评框架。[Read In English][paper]
[GAOKAO-MM]:基于中国高考题的多模态数据集,测评多模态模型的感知、理解、知识、推理能力。
[GAOKAO-Bench-2023]:将中国2023年高考选择题作为数据集 ,对GAOKAO-Bench的补充。
我们希望能够建立一个标准化、综合性的评测框架来对大模型进行全方位、准确的评估。在中国,高考是标准化水平最高、综合性最强并且认可度最广的考试之一,我们希望借用高考的题目来评估大模型的能力。因此,我们收集了2010-2022年全国高考卷的题目,其中包括1781道客观题和1030道主观题,构建起GAOKAO-Bench的数据部分。
题目类型 | 题目数量 | 数量占比 |
---|---|---|
客观题 | 1781 | 63.36% |
主观题 | 1030 | 36.64% |
题目总数 | 2811 | 100% |
数据示例如下所示:
- Year
2022
- Category
全国甲卷
- Score
5
- Question
若
$z=-1+\sqrt{3} \mathrm{i}$ , 则$\frac{z}{z \bar{z}-1}=()$ A.
$-1+\sqrt{3} \mathrm{i}$ B.
$-1-\sqrt{3} i$ C.
$-\frac{1}{3}+\frac{\sqrt{3}}{3} \mathrm{i}$ D.
$-\frac{1}{3}-\frac{\sqrt{3}}{3} i$
- Analysis
【详解】
$\bar{z}=-1-\sqrt{3} i, z \bar{z}=(-1+\sqrt{3} i)(-1-\sqrt{3} i)=1+3=4$ .
$\frac{z}{z \bar{z}-1}=\frac{-1+\sqrt{3} \mathrm{i}}{3}=-\frac{1}{3}+\frac{\sqrt{3}}{3} \mathrm{i}$ 故选: C
- Standard Answer
C
我们采用zero-shot的方式测试各项模型,对客观题采用基于规则的答案抽取方式,对主观题采取人工评阅的方式,最终获得了GPT-4、GPT-3.5等模型的转化后的高考总分。实验结果表明,GPT-4转换后的高考总分名列第一,文科和理科总分分别为485和447。同时,所有模型的文科成绩都高于理科成绩。
Models | Overall | Chinese | Eng. | Sci. Math | Hum. Math | Phys. | Chem. | Biol. | Poli. | Hist. | Geog. |
---|---|---|---|---|---|---|---|---|---|---|---|
GPT-4-0314 | 72.2% | 53.9% | 93.1% | 53.7% | 63.3% | 55.5% | 44.4% | 80.7% | 75.9% | 75.6% | 80.0% |
GPT-4-0613 | 71.6% | 52.1% | 93.2% | 54.5% | 64.0% | 50.8% | 43.6% | 83.0% | 72.5% | 74.2% | 81.1% |
Gemini-Pro | 57.9% | 46.7% | 69.9% | 40.7% | 47.7% | 32.0% | 40.3% | 70.7% | 64.7% | 64.5% | 68.4% |
ERNIE-Bot-0615 | 56.6% | 46.7% | 31.0% | 38.3% | 49.1% | 35.9% | 66.1% | 79.3% | 86.9% | 79.1% | 68.4% |
GPT-3.5-turbo-0301 | 53.2% | 34.7% | 76.6% | 38.8% | 47.8% | 41.1% | 38.7% | 56.9% | 45.3% | 53.9% | 54.0% |
ERNIE-Bot-turbo-0725 | 45.6% | 35.3% | 26.6% | 34.1% | 36.2% | 32.0% | 51.6% | 64.0% | 72.2% | 63.4% | 44.2% |
Baichuan2-13b-Chat | 43.9% | 26.9% | 34.7% | 23.8% | 31.7% | 25.0% | 40.3% | 53.3% | 75.3% | 59.9% | 61.1% |
ChatGLM2-6b | 42.7% | 31.1% | 30.6% | 29.0% | 35.8% | 24.2% | 46.0% | 71.3% | 55.0% | 59.2% | 41.1% |
Baichuan2-7b-Chat | 40.5% | 31.7% | 33.0% | 26.6% | 28.4% | 18.0% | 26.6% | 48.0% | 69.7% | 57.8% | 49.5% |
ChatGLM-6b | 30.8% | 18.6% | 17.0% | 25.2% | 25.7% | 12.5% | 30.6% | 24.7% | 54.1% | 59.9% | 25.3% |
Baichuan2-7b-Base | 27.2% | 16.2% | 21.2% | 24.8% | 24.8% | 0.0% | 23.4% | 24.0% | 55.3% | 32.1% | 24.2% |
LLaMA-7b | 21.1% | 16.2% | 20.5% | 24.3% | 26.1% | 0.0% | 22.6% | 22.7% | 22.2% | 19.2% | 24.2% |
Vicuna-7b | 21.0% | 12.0% | 19.6% | 23.8% | 23.4% | 7.0% | 27.4% | 20.0% | 20.9% | 23.0% | 23.2% |
Models | Overall | Chinese | Eng. | Sci. Math | Hum. Math | Phys. | Chem. | Biol. | Poli. | Hist. | Geog. |
---|---|---|---|---|---|---|---|---|---|---|---|
GPT-4-0314 | 51.9% | 51.5% | 88.3% | 24.1% | 27.9% | 56.7% | 35.0% | 85.6% | 50.0% | 63.1% | 70.0% |
GPT-4-0613 | 50.8% | 50.3% | 87.6% | 24.6% | 27.5% | 47.1% | 28.5% | 85.6% | 49.9% | 59.9% | 71.5% |
ERNIE-Bot-0615 | 48.4% | 57.1% | 45.0% | 17.0% | 25.6% | 33.5% | 30.8% | 84.9% | 53.0% | 60.0% | 72.7% |
ERNIE-Bot-turbo-0725 | 39.2% | 42.5% | 28.8% | 14.6% | 15.6% | 23.2% | 25.0% | 85.1% | 45.3% | 47.0% | 61.8% |
GPT-3.5-turbo-0301 | 35.8% | 33.9% | 75.4% | 15.2% | 15.9% | 16.9% | 21.4% | 36.3% | 42.3% | 58.4% | 62.1% |
-
获取GPT-4模型输出
cd ./Bench ## Get the Output of Objective Questions python objective_bench.py --openai_api_key="your openai api key" ## Get the Output of Subjective Questions python subjective_bench.py --openai_api_key="your openai api key"
-
计算GPT-4模型客观题得分率
-
将GPT-4模型输出的JSON文件存放在
./Results/gpt_4_obj
文件夹下。 -
执行以下指令,获得其客观题的得分率,结果存放在
./Results/gpt_4_obj/result/correction_score.json
文件下。
python OBJ_score_evaluation.py --obj_output_dir=../Results/gpt_4_obj
-
-
计算GPT-4模型主观题得分率
由于人工批改的高昂成本,我们提供了LLM-as-a-Judge脚本,利用GPT-4-turbo为模型的主观题打分。
-
将GPT-4模型输出的JSON文件存放在
./Results/gpt_4_sub
文件夹下。 -
执行以下指令,获得GPT-4对主观题的评分,结果存放在
./Results/gpt_4_sub/gpt-4-1106-preview_correction_wo_marking_criterion
文件下。
python subjective_grade.py --openai_api_key="your openai api key"
-
执行以下指令,获得其主观题的得分率,结果存放在
./Results/gpt_4_sub/gpt-4-1106-preview_correction_wo_marking_criterion/result/model_score.json
文件下。python SUB_score_evaluation.py --sub_output_dir=../Results/gpt_4_sub/gpt-4-1106-preview_correction_wo_marking_criterion --mode=model
-
-
计算GPT-4模型高考总分
执行以下指令,获得GPT-4转换后的高考总分,结果保存在
./Results/merge_score.json
下。python merge_OBJ_SUB_score.py
封装你的模型API并放置在 ./Models
目录下,封装方式可参考./Models/openai_gpt4.py
。
@inproceedings{Zhang2023EvaluatingTP,
title={Evaluating the Performance of Large Language Models on GAOKAO Benchmark},
author={Xiaotian Zhang and Chunyang Li and Yi Zong and Zhengyu Ying and Liang He and Xipeng Qiu},
year={2023}
}
我们非常感谢上海市曹杨第二中学的老师们,他们负责了GAOKAO-Bench主观题部分的评分。