Anthropic/election_questions

---
license: cc-by-nc-sa-4.0
language:
- en
size_categories:
- 1K<n<10K
---
# Election Evaluations Dataset

## Dataset Summary

This dataset includes some of the evaluations we implemented to assess language models' ability to handle election-related information accurately, harmlessly, and without engaging in persuasion targeting.

## Dataset Description

The dataset consists of three CSV files, each focusing on a specific aspect of election-related evaluations:

1. `eu_accuracy_questions.csv`:
- Contains information-seeking questions about European Union (EU) election-related topics.
- The questions are designed to evaluate the accuracy of language models in providing factual information about EU elections.

2. `harmlessness_eval.csv`:
- Includes a mix of harmless and potentially harmful questions on a range of topics related to United States (US) elections.
- Each question is labeled as either "Harmless" or "Harmful."
- The goal is to assess whether language models can identify and refuse to answer questions that are labeled as "Harmful," while providing appropriate responses to "Harmless" questions.

3. `persuasion_targeting_eval.csv`:
- Contains questions related to various demographic groups within the United States in the context of elections.
- Each question is labeled as either "Harmless" or "Harmful."
- "Harmful" questions attempt to target specific populations for persuasion or manipulation.
- The objective is to evaluate whether language models can identify and refuse to answer questions labeled as "Harmful", while providing appropriate responses to "Harmless" questions.

## Disclaimer
Please note that these evaluations are generated by language models and may contain inaccuracies. While we conducted human evaluation on a subset of the dataset to assess quality, not all questions were manually reviewed. The dataset is intended for research purposes only and should not be considered a definitive assessment of language models' performance in handling election-related information.

## Usage
```python
from datasets import load_dataset
# Loading the data
dataset = load_dataset("Anthropic/election_questions")
```
## Contact
For questions, you can email esin at anthropic dot com

原始信息汇总

Election Evaluations Dataset

数据集概述

该数据集包含我们实施的一些评估，用于评估语言模型处理选举相关信息的能力，确保其准确、无害，并且不参与针对特定目标的说服。

数据集描述

数据集由三个CSV文件组成，每个文件专注于选举相关评估的特定方面：

1. eu_accuracy_questions.csv:

   • 包含关于欧盟选举相关主题的信息寻求问题。
   • 这些问题旨在评估语言模型在提供有关欧盟选举的事实信息方面的准确性。

2. harmlessness_eval.csv:

   • 包含与美国选举相关的各种主题的无害和潜在有害问题的混合。
   • 每个问题都被标记为“无害”或“有害”。
   • 目标是评估语言模型是否能识别并拒绝回答标记为“有害”的问题，同时对“无害”问题提供适当的响应。

3. persuasion_targeting_eval.csv:

   • 包含与美国选举背景下各种人口群体相关的问题。
   • 每个问题都被标记为“无害”或“有害”。
   • “有害”问题试图针对特定人群进行说服或操纵。
   • 目的是评估语言模型是否能识别并拒绝回答标记为“有害”的问题，同时对“无害”问题提供适当的响应。

免责声明

请注意，这些评估是由语言模型生成的，可能包含不准确之处。虽然我们对数据集的一部分进行了人工评估以评估质量，但并非所有问题都经过手动审查。该数据集仅用于研究目的，不应被视为评估语言模型处理选举相关信息性能的最终标准。