values-in-the-wild

## Summary
This dataset presents a comprehensive taxonomy of 3307 values expressed by Claude (an AI assistant) across hundreds of thousands of real-world conversations. Using a novel privacy-preserving methodology, these values were extracted and classified without human reviewers accessing any conversation content. The dataset reveals patterns in how AI systems express values "in the wild" when interacting with diverse users and tasks.

We're releasing this resource to advance research in two key areas: understanding value expression in deployed language models and supporting broader values research across disciplines. By providing empirical data on AI values "in the wild," we hope to move toward a more grounded understanding of how values manifest in human-AI interactions.

For information on how this dataset was constructed, and related analysis, please see the accompanying paper: [Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions](https://assets.anthropic.com/m/18d20cca3cde3503/original/Values-in-the-Wild-Paper.pdf).

**Note:** You can interpret the occurrence of each value in the dataset as "The AI's response demonstrated valuing {VALUE}." For example, for the value of "accuracy" (5.3% frequency), this means that our methods detected that Claude's response demonstrated *valuing* accuracy 5.3% of the time (not that it *was* accurate in 5.3% of conversations).

## Dataset Description

The dataset includes two CSV files:

1. `values_frequencies.csv`
- This shows every extracted AI value along with their frequency of occurrence across the conversation sample. There are two columns:
- `value`: The value label (e.g. `accuracy` or `helpfulness`).
- `pct_convos`: The percentage of the subjective conversation sample that that this value was detected in, rounded to 3 decimal places.
- This is sorted by the `pct_convos` column.
2. `values_tree.csv`
- This shows the hierarchical taxonomy of values, where we sequentially cluster/group the values into higher-level categories. There are six columns:
- `cluster_id`: If `level > 0`, this denotes the ID of the cluster of values. If `level = 0`, this is just identical to the `name` of the extracted value.
- `description`: If `level > 0`, the Claude-generated description of the cluster of values.
- `name`: The name of the extracted value itself (if `level = 0`, or the cluster of values (if `level > 0`).
- `level`: Out of `0, 1, 2, 3`, which level of the taxonomy is this value/cluster of values at. `level = 0` means the lowest level, i.e. the individual values; `level = 3` is the highest level (e.g. "Epistemic values").
- `parent_cluster_id`: The `cluster_id` of the higher-level parent cluster of this.
- `pct_total_occurrences`: The percentage of the total *number of values expressions* that was expressions of this value, rounded to 3 decimal places.
- This is sorted by the `parent_cluster_id` column, so that values clustered together appear together.

## Disclaimer
Please note that the extracted values, descriptions and cluster names were generated by a language model and may contain inaccuracies. While we conducted human evaluation on our values extractor to assess quality, and manually checked the hierarchy for clarity and accuracy, inferring values is an inherently subjective endeavor, and there may still be errors. The dataset is intended for research purposes only and should not be considered a definitive assessment of what values may be expressed by Claude, or language models in general.

## Usage
```python
from datasets import load_dataset
dataset_values_frequencies = load_dataset("Anthropic/values-in-the-wild", "values_frequencies")
dataset_values_tree = load_dataset("Anthropic/values-in-the-wild", "values_tree")
```

## Contact
For questions, you can email saffron at anthropic dot com

## 摘要
本数据集呈现了一套完整的价值分类体系，涵盖了AI助手Claude在数十万次真实对话中所体现的3307项价值主张。本研究采用了一种新型隐私保护方法，在无需人工审阅者访问任何对话内容的前提下，完成了价值主张的提取与分类。该数据集揭示了AI系统在与多样化用户及各类任务交互时，于真实场景中展现价值倾向的模式。

我们发布此资源旨在推动两大关键领域的研究：一是理解部署的大语言模型（Large Language Model）中的价值表达，二是支持跨学科的广义价值研究。通过提供AI价值在真实场景下的实证数据，我们期望推动对人类与AI交互中价值显现方式的更具实证基础的理解。

如需了解本数据集的构建方式及相关分析，请参阅配套论文：《野外的价值：在真实世界语言模型交互中发掘与分析价值主张》（Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions），链接为：https://assets.anthropic.com/m/18d20cca3cde3503/original/Values-in-the-Wild-Paper.pdf。

**注：** 你可将数据集中每项价值的出现解读为“AI的响应体现了对{VALUE}的重视”。例如，针对“准确性”（出现频率为5.3%）这一价值，这意味着我们的方法检测到Claude的响应在5.3%的场景中体现了对准确性的重视（而非5.3%的对话中其输出准确无误）。

## 数据集说明
本数据集包含两个CSV文件：
1. `values_frequencies.csv`
该文件列出了所有提取出的AI价值主张及其在对话样本中的出现频率。包含两列：
- `value`：价值标签（例如`accuracy`或`helpfulness`）。
- `pct_convos`：该价值被检测到的对话样本占比，保留三位小数。
本文件按`pct_convos`列排序。
2. `values_tree.csv`
该文件展示了价值的层级分类体系，即将各项价值依次聚类为更高层级的类别。包含六列：
- `cluster_id`：若`level > 0`，则表示该值聚类的ID；若`level = 0`，则与提取出的价值的`name`字段完全一致。
- `description`：若`level > 0`，则为Claude生成的该值聚类的描述。
- `name`：提取出的单个价值的名称（若`level = 0`），或该值聚类的名称（若`level > 0`）。
- `level`：取值为`0, 1, 2, 3`，表示该价值/值聚类在分类体系中的层级。`level = 0`代表最低层级，即单个价值；`level = 3`为最高层级（例如“认知价值”）。
- `parent_cluster_id`：该值聚类的上级父聚类的`cluster_id`。
- `pct_total_occurrences`：该价值/值聚类的表达占所有价值表达总数的百分比，保留三位小数。
本文件按`parent_cluster_id`列排序，因此同属一类的价值会集中呈现。

## 免责声明
请注意，提取出的价值主张、描述及聚类名称均由语言模型生成，可能存在不准确之处。尽管我们已对价值提取器开展了人工评估以验证其质量，并手动核查了分类体系的清晰性与准确性，但价值推断本质上是一项主观工作，仍可能存在错误。本数据集仅用于研究目的，不应被视为对Claude或一般语言模型所可能体现的价值的确定性评估。

## 使用方法
python
from datasets import load_dataset
dataset_values_frequencies = load_dataset("Anthropic/values-in-the-wild", "values_frequencies")
dataset_values_tree = load_dataset("Anthropic/values-in-the-wild", "values_tree")

## 联系方式
如有疑问，请发送邮件至[e-mail]。

搜集汇总

背景与挑战

背景概述

该数据集通过隐私保护方法，从Claude AI助手在大量真实对话中提取并分类了3307个表达值，构建了一个层次化分类体系。它旨在为研究AI系统在部署中的价值表达以及跨学科价值研究提供实证数据，包含两个CSV文件分别记录值频率和分类树结构。