hendrycks/ethics
收藏资源简介:
--- license: mit language: en dataset_info: - config_name: default features: - name: label dtype: int64 - name: input dtype: string - config_name: commonsense features: - name: label dtype: int32 - name: input dtype: string splits: - name: train num_bytes: 14429921 num_examples: 13910 - name: validation num_bytes: 3148616 num_examples: 3885 - name: test num_bytes: 3863068 num_examples: 3964 download_size: 21625153 dataset_size: 21441605 - config_name: deontology features: - name: label dtype: int32 - name: scenario dtype: string - name: excuse dtype: string splits: - name: train num_bytes: 1854277 num_examples: 18164 - name: validation num_bytes: 369318 num_examples: 3596 - name: test num_bytes: 359268 num_examples: 3536 download_size: 2384007 dataset_size: 2582863 - config_name: justice features: - name: label dtype: int32 - name: scenario dtype: string splits: - name: train num_bytes: 2423889 num_examples: 21791 - name: validation num_bytes: 297935 num_examples: 2704 - name: test num_bytes: 228008 num_examples: 2052 download_size: 2837375 dataset_size: 2949832 - config_name: utilitarianism features: - name: baseline dtype: string - name: less_pleasant dtype: string splits: - name: train num_bytes: 2186713 num_examples: 13737 - name: validation num_bytes: 730391 num_examples: 4807 - name: test num_bytes: 668429 num_examples: 4271 download_size: 3466564 dataset_size: 3585533 - config_name: virtue features: - name: label dtype: int32 - name: scenario dtype: string splits: - name: train num_bytes: 2605021 num_examples: 28245 - name: validation num_bytes: 467254 num_examples: 4975 - name: test num_bytes: 452491 num_examples: 4780 download_size: 3364070 dataset_size: 3524766 tags: - AI Alignment --- # Dataset Card for ETHICS This is the data from [Aligning AI With Shared Human Values](https://arxiv.org/pdf/2008.02275) by Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt, published at ICLR 2021. For more information, see the [Github Repo](https://github.com/hendrycks/ethics). ## Dataset Summary This dataset provides ethics-based tasks for evaluating language models for AI alignment. ## Loading Data To load this data, you can use HuggingFace datasets and the dataloader script. ``` from datasets import load_dataset load_dataset("hendrycks/ethics", "commonsense") ``` Where `commonsense` is one of the following sections: commonsense, deontology, justice, utilitarianism, and virtue. ### Citation Information ``` @article{hendrycks2021ethics, title={Aligning AI With Shared Human Values}, author={Dan Hendrycks and Collin Burns and Steven Basart and Andrew Critch and Jerry Li and Dawn Song and Jacob Steinhardt}, journal={Proceedings of the International Conference on Learning Representations (ICLR)}, year={2021} } ```
license: MIT许可证 language: 英语 dataset_info: - config_name: default(默认配置) features: - 名称: label 数据类型: int64 - 名称: input 数据类型: 字符串 - config_name: commonsense(常识伦理配置) features: - 名称: label 数据类型: int32 - 名称: input 数据类型: 字符串 splits: - 拆分集名称: 训练集(train) 字节数: 14429921 样本数量: 13910 - 拆分集名称: 验证集(validation) 字节数: 3148616 样本数量: 3885 - 拆分集名称: 测试集(test) 字节数: 3863068 样本数量: 3964 下载大小: 21625153 数据集总大小: 21441605 - config_name: deontology(义务论,deontology) features: - 名称: label 数据类型: int32 - 名称: scenario 数据类型: 字符串(场景) - 名称: excuse 数据类型: 字符串(托词) splits: - 拆分集名称: 训练集(train) 字节数: 1854277 样本数量: 18164 - 拆分集名称: 验证集(validation) 字节数: 369318 样本数量: 3596 - 拆分集名称: 测试集(test) 字节数: 359268 样本数量: 3536 下载大小: 2384007 数据集总大小: 2582863 - config_name: justice(正义论,justice) features: - 名称: label 数据类型: int32 - 名称: scenario 数据类型: 字符串(场景) splits: - 拆分集名称: 训练集(train) 字节数: 2423889 样本数量: 21791 - 拆分集名称: 验证集(validation) 字节数: 297935 样本数量: 2704 - 拆分集名称: 测试集(test) 字节数: 228008 样本数量: 2052 下载大小: 2837375 数据集总大小: 2949832 - config_name: utilitarianism(功利主义,utilitarianism) features: - 名称: baseline 数据类型: 字符串(基准方案) - 名称: less_pleasant 数据类型: 字符串(较不愉悦选项) splits: - 拆分集名称: 训练集(train) 字节数: 2186713 样本数量: 13737 - 拆分集名称: 验证集(validation) 字节数: 730391 样本数量: 4807 - 拆分集名称: 测试集(test) 字节数: 668429 样本数量: 4271 下载大小: 3466564 数据集总大小: 3585533 - config_name: virtue(美德伦理学,virtue) features: - 名称: label 数据类型: int32 - 名称: scenario 数据类型: 字符串(场景) splits: - 拆分集名称: 训练集(train) 字节数: 2605021 样本数量: 28245 - 拆分集名称: 验证集(validation) 字节数: 467254 样本数量: 4975 - 拆分集名称: 测试集(test) 字节数: 452491 样本数量: 4780 下载大小: 3364070 数据集总大小: 3524766 tags: - 人工智能对齐(AI Alignment) # ETHICS数据集卡片 本数据集源自Dan Hendrycks、Collin Burns、Steven Basart、Andrew Critch、Jerry Li、Dawn Song与Jacob Steinhardt合著的论文《Aligning AI With Shared Human Values》,该论文发表于2021年国际学习表征会议(ICLR),原文链接:https://arxiv.org/pdf/2008.02275。 更多详细信息可访问其GitHub仓库:https://github.com/hendrycks/ethics。 ## 数据集概述 本数据集包含一系列基于伦理学的任务,用于评估面向人工智能对齐(AI Alignment)的大语言模型。 ## 数据加载方法 可通过HuggingFace Datasets库及对应数据加载脚本加载本数据集,示例代码如下: python from datasets import load_dataset load_dataset("hendrycks/ethics", "commonsense") 其中`commonsense`为下述配置名称之一:常识伦理、义务论、正义论、功利主义与美德伦理学。 ### 引用信息 bibtex @article{hendrycks2021ethics, title={Aligning AI With Shared Human Values}, author={Dan Hendrycks and Collin Burns and Steven Basart and Andrew Critch and Jerry Li and Dawn Song and Jacob Steinhardt}, journal={Proceedings of the International Conference on Learning Representations (ICLR)}, year={2021} }
数据集概述
数据集配置
默认配置
- 特征:
label: 类型int64input: 类型string
常识配置
- 特征:
label: 类型int32input: 类型string
- 分割:
train: 字节数14429921, 样本数13910validation: 字节数3148616, 样本数3885test: 字节数3863068, 样本数3964
- 下载大小:
21625153字节 - 数据集大小:
21441605字节
道义论配置
- 特征:
label: 类型int32scenario: 类型stringexcuse: 类型string
- 分割:
train: 字节数1854277, 样本数18164validation: 字节数369318, 样本数3596test: 字节数359268, 样本数3536
- 下载大小:
2384007字节 - 数据集大小:
2582863字节
正义配置
- 特征:
label: 类型int32scenario: 类型string
- 分割:
train: 字节数2423889, 样本数21791validation: 字节数297935, 样本数2704test: 字节数228008, 样本数2052
- 下载大小:
2837375字节 - 数据集大小:
2949832字节
功利主义配置
- 特征:
baseline: 类型stringless_pleasant: 类型string
- 分割:
train: 字节数2186713, 样本数13737validation: 字节数730391, 样本数4807test: 字节数668429, 样本数4271
- 下载大小:
3466564字节 - 数据集大小:
3585533字节
美德配置
- 特征:
label: 类型int32scenario: 类型string
- 分割:
train: 字节数2605021, 样本数28245validation: 字节数467254, 样本数4975test: 字节数452491, 样本数4780
- 下载大小:
3364070字节 - 数据集大小:
3524766字节




