SeaLLMs/SeaExam
收藏资源简介:
--- license: apache-2.0 configs: - config_name: m3exam-chinese data_files: - split: dev path: m3exam-chinese/dev.json - split: test path: m3exam-chinese/test.json - config_name: m3exam-english data_files: - split: dev path: m3exam-english/dev.json - split: test path: m3exam-english/test.json - config_name: m3exam-thai data_files: - split: dev path: m3exam-thai/dev.json - split: test path: m3exam-thai/test.json - config_name: m3exam-vietnamese data_files: - split: dev path: m3exam-vietnamese/dev.json - split: test path: m3exam-vietnamese/test.json - config_name: m3exam-indonesian data_files: - split: dev path: m3exam-indonesian/dev.json - split: test path: m3exam-indonesian/test.json - config_name: mmlu-english data_files: - split: dev path: mmlu-english/dev.json - split: test path: mmlu-english/test.json - config_name: mmlu-chinese data_files: - split: dev path: mmlu-chinese/dev.json - split: test path: mmlu-chinese/test.json - config_name: mmlu-thai data_files: - split: dev path: mmlu-thai/dev.json - split: test path: mmlu-thai/test.json - config_name: mmlu-vietnamese data_files: - split: dev path: mmlu-vietnamese/dev.json - split: test path: mmlu-vietnamese/test.json - config_name: mmlu-indonesian data_files: - split: dev path: mmlu-indonesian/dev.json - split: test path: mmlu-indonesian/test.json task_categories: - multiple-choice language: - en - id - vi - th - zh tags: - exam --- > Check the 🏆 [leaderboard](https://huggingface.co/spaces/SeaLLMs/SeaExam_leaderboard) constructed with this dataset and the corresponding 👨🏻💻 [evaluation code](https://github.com/DAMO-NLP-SG/SeaExam). # SeaExam dataset The SeaExam dataset aims to evaluate Large Language Models (LLMs) on a diverse set of Southeast Asian (SEA) languages including English, Chinese, Indonesian, Thai, and Vietnamese. Our goal is to ensure a fair and consistent comparison across different LLMs on those languages while mitigating the risk of data contamination. It consists of the following two parts: ### M3Exam (with adjustments) The original [M3Exam](https://github.com/DAMO-NLP-SG/M3Exam) dataset is constructed with real human exam questions collected from different countries. As a result, the dataset retains the diverse cultural characteristics inherent in the questions. We further process the original dataset with the following operations: - We standardized the total number of answer options to four. This involved removing questions with fewer than four options and eliminating one incorrect option from questions that initially had more than four options. - All answers have been mapped to a numerical value within the range [0, 1, 2, 3] for consistency. - We removed the option index from each answer choice (e.g., changing "A. good" to "good") to simplify the format. - Randomly shuffle the options. ### Translated MMLU The [MMLU](https://github.com/hendrycks/test) dataset contains English questions from 57 subjects. We translate the original English questions to different languages to measure the cross-lingual alignment: - We randomly selected 50 questions from each subject, totaling 2850 questions. - These questions have been translated from English into Chinese, Indonesian, Thai, and Vietnamese using Google Translate to ensure linguistic diversity. - Randomly shuffle the options. # Usage To load a particular subset of the dataset, you need to specify the sub-dataset name of the language. For example, ```python from datasets import load_dataset ds_name = "m3exam" lang = "english" dataset = load_dataset(f"SeaLLMs/SeaExam", f"{ds_name}-{lang}") ``` To load the whole dataset: ```python from datasets import load_dataset for ds_name in ['m3exam','mmlu']: for lang in ['english', 'chinese', 'thai', 'vietnamese', 'indonesian']: dataset = load_dataset(f"SeaLLMs/SeaExam", f"{ds_name}-{lang}") print(dataset) ```
许可证:Apache-2.0 配置项: - 配置名称:m3exam-chinese 数据文件: - 划分集:验证集,路径:m3exam-chinese/dev.json - 划分集:测试集,路径:m3exam-chinese/test.json - 配置名称:m3exam-english 数据文件: - 划分集:验证集,路径:m3exam-english/dev.json - 划分集:测试集,路径:m3exam-english/test.json - 配置名称:m3exam-thai 数据文件: - 划分集:验证集,路径:m3exam-thai/dev.json - 划分集:测试集,路径:m3exam-thai/test.json - 配置名称:m3exam-vietnamese 数据文件: - 划分集:验证集,路径:m3exam-vietnamese/dev.json - 划分集:测试集,路径:m3exam-vietnamese/test.json - 配置名称:m3exam-indonesian 数据文件: - 划分集:验证集,路径:m3exam-indonesian/dev.json - 划分集:测试集,路径:m3exam-indonesian/test.json - 配置名称:mmlu-english 数据文件: - 划分集:验证集,路径:mmlu-english/dev.json - 划分集:测试集,路径:mmlu-english/test.json - 配置名称:mmlu-chinese 数据文件: - 划分集:验证集,路径:mmlu-chinese/dev.json - 划分集:测试集,路径:mmlu-chinese/test.json - 配置名称:mmlu-thai 数据文件: - 划分集:验证集,路径:mmlu-thai/dev.json - 划分集:测试集,路径:mmlu-thai/test.json - 配置名称:mmlu-vietnamese 数据文件: - 划分集:验证集,路径:mmlu-vietnamese/dev.json - 划分集:测试集,路径:mmlu-vietnamese/test.json - 配置名称:mmlu-indonesian 数据文件: - 划分集:验证集,路径:mmlu-indonesian/dev.json - 划分集:测试集,路径:mmlu-indonesian/test.json 任务类别:多项选择 语言:英语、印尼语、越南语、泰语、中文 标签:考试 > 可查看基于本数据集构建的🏆[排行榜](https://huggingface.co/spaces/SeaLLMs/SeaExam_leaderboard)以及对应的👨🏻💻[评估代码](https://github.com/DAMO-NLP-SG/SeaExam)。 # SeaExam 数据集 SeaExam 数据集旨在针对涵盖英语、中文、印尼语、泰语及越南语的多语种东南亚(SEA)语言集合,对大语言模型(Large Language Model, LLM)进行性能评估。本数据集的目标是确保不同大语言模型在上述语言上的评估结果具备公平性与一致性,同时降低数据污染风险。本数据集包含以下两个部分: ### 经过调整的M3考试(M3Exam)数据集 原始[M3考试(M3Exam)](https://github.com/DAMO-NLP-SG/M3Exam)数据集由从多国收集的真实人类考试试题构建而成,因此保留了试题中蕴含的多元文化特征。我们对原始数据集进行了如下预处理操作: - 将答案选项的总数统一为四项:移除选项数不足四项的试题,并从初始选项数超过四项的试题中剔除一个错误选项。 - 为保证格式统一,我们将所有答案映射至[0, 1, 2, 3]的数值范围内。 - 移除每个答案选项前的索引标识(例如将"A. good"调整为"good")以简化格式。 - 对选项进行随机打乱。 ### 翻译版MMLU数据集 [MMLU](https://github.com/hendrycks/test)数据集包含来自57个学科的英语试题。为了评估跨语言对齐能力,我们将原始英语试题翻译为其他语言: - 从每个学科中随机选取50道试题,总计2850道。 - 为保证语言多样性,我们通过谷歌翻译将这些试题从英语翻译至中文、印尼语、泰语及越南语。 - 对选项进行随机打乱。 # 使用方法 若需加载数据集的特定子集,需指定对应语言的子数据集名称,示例如下: python from datasets import load_dataset ds_name = "m3exam" lang = "english" dataset = load_dataset(f"SeaLLMs/SeaExam", f"{ds_name}-{lang}") 若需加载完整数据集: python from datasets import load_dataset for ds_name in ['m3exam','mmlu']: for lang in ['english', 'chinese', 'thai', 'vietnamese', 'indonesian']: dataset = load_dataset(f"SeaLLMs/SeaExam", f"{ds_name}-{lang}") print(dataset)
SeaExam 数据集
SeaExam 数据集旨在评估大型语言模型(LLMs)在包括英语、中文、印尼语、泰语和越南语在内的多种东南亚语言上的表现。数据集包括以下两部分:
M3Exam(调整版)
原始的 M3Exam 数据集由来自不同国家的真实人类考试题目组成,保留了问题中固有的多样文化特征。我们对原始数据集进行了以下处理:
- 标准化答案选项总数为四个,移除了少于四个选项的问题,并从最初有超过四个选项的问题中删除一个错误选项。
- 所有答案已映射到 [0, 1, 2, 3] 范围内的数值,以保持一致性。
- 移除了每个答案选项的索引(例如,将 "A. good" 改为 "good")以简化格式。
- 随机打乱选项。
翻译的 MMLU
MMLU 数据集包含来自 57 个科目的英语问题。我们将原始英语问题翻译成不同语言以衡量跨语言对齐:
- 从每个科目中随机选择了 50 个问题,总计 2850 个问题。
- 这些问题已从英语翻译成中文、印尼语、泰语和越南语,使用 Google Translate 以确保语言多样性。
- 随机打乱选项。
使用方法
要加载特定子集的数据集,需要指定语言的子数据集名称。例如: python from datasets import load_dataset
ds_name = "m3exam" lang = "english" dataset = load_dataset(f"SeaLLMs/SeaExam", f"{ds_name}-{lang}")
要加载整个数据集: python from datasets import load_dataset
for ds_name in [m3exam,mmlu]: for lang in [english, chinese, thai, vietnamese, indonesian]: dataset = load_dataset(f"SeaLLMs/SeaExam", f"{ds_name}-{lang}") print(dataset)




