https://github.com/aygp-dr/values-compass

Tools for exploring and analyzing Anthropic's Values-in-the-Wild dataset for AI ethics research
https://github.com/aygp-dr/values-compass

ai-ethics anthropic-claude data-analysis nlp values

Last synced: 5 months ago
JSON representation

Tools for exploring and analyzing Anthropic's Values-in-the-Wild dataset for AI ethics research

• Host: GitHub
• URL: https://github.com/aygp-dr/values-compass
• Owner: aygp-dr
• Created: 2025-04-30T07:56:45.000Z (5 months ago)
• Default Branch: main
• Last Pushed: 2025-05-04T05:21:17.000Z (5 months ago)
• Last Synced: 2025-05-12T21:59:42.911Z (5 months ago)
• Topics: ai-ethics, anthropic-claude, data-analysis, nlp, values
• Language: Python
• Size: 3.88 MB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 2
• Metadata Files:
  • Readme: README.org

Awesome Lists containing this project

README

          
#+TITLE: Exploring Anthropic's Values-in-the-Wild Dataset

#+AUTHOR: Aidan Pace

#+EMAIL: apace@defrecord.com

#+DATE: [2025-04-30 Wed]

* Values in the Wild: Understanding AI Values System

This project provides tools for exploring and analyzing Anthropic's Values-in-the-Wild dataset, which offers a comprehensive taxonomy of values expressed by AI assistants in real-world interactions.

** Background

The Values-in-the-Wild dataset is a groundbreaking resource from Anthropic that catalogs and taxonomizes 3,307 unique values expressed by Claude during hundreds of thousands of real-world conversations. Using a privacy-preserving methodology, Anthropic researchers extracted these values without human reviewers needing to access any conversation content.

The research presents the first large-scale empirical taxonomy of AI values observed in deployment contexts, revealing both consistent patterns and contextual variations in how AI systems express values across different interaction types.

** Dataset Overview

The dataset contains two main configurations:

- ~values_frequencies~: Contains data about how frequently different values appear

- ~values_tree~: Contains the hierarchical structure of the values taxonomy

The values are organized into five primary conceptual domains:

1. Practical Values (31.4%)

2. Epistemic Values (22.2%)

3. Social Values (21.4%)

4. Protective Values (13.9%)

5. Personal Values (11.1%)

The most common AI values include:

- Helpfulness (23.4%)

- Professionalism (22.9%)

- Transparency (17.4%)

- Clarity (16.6%)

- Thoroughness (14.3%)

** Key Findings

The research demonstrates that while Claude expresses thousands of diverse values that respond to and engage with varied human perspectives, it tends to express some common trans-situational values—primarily centered around competent and supportive assistance, such as "helpfulness," "professionalism," "transparency," and "clarity."

Values are highly context-dependent:

- "Healthy boundaries" appears disproportionately when providing relationship advice

- "Historical accuracy" appears when analyzing controversial historical events

- "Human agency" features prominently in technology ethics discussions

Claude's response to human values varies:

- It often mirrors positive human values (e.g., responding to "authenticity" with "authenticity")

- It counters values like "deception" with "ethical integrity" and "honesty"

** References

1. Anthropic. (2025). Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions. [[https://www.anthropic.com/research/values-wild][Anthropic Research]]

2. Huang, S., Durmus, E., McCain, M., Handa, K., Tamkin, A., Hong, J., Stern, M., Somani, A., Zhang, X., & Ganguli, D. (2025). 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][Research Paper]]

3. Hoover, S. (2025). Anthropic mapped Claude's morality: Here's what the chatbot values and doesn't. [[https://www.zdnet.com/article/anthropic-mapped-claudes-morality-heres-what-the-chatbot-values-and-doesnt/][ZDNet]]

4. Borges, J.L. (1941). "The Library of Babel," in *The Garden of Forking Paths*.

5. Saunders, G. (2021). "A Swim in a Pond in the Rain: In Which Four Russians Give a Master Class on Writing, Reading, and Life."

** Dataset Access

The dataset is available on Hugging Face:

#+BEGIN_SRC 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")

#+END_SRC

** Project Structure

This repository contains tools for exploring and analyzing the Values-in-the-Wild dataset:

#+BEGIN_SRC

values_explorer/

├── .gitignore

├── README.org

├── requirements.txt

├── setup.py

├── notebooks/

│   └── exploration.org

├── stories/                                   # Creative explorations (not primary analysis)

│   ├── README.org

│   └── images/

└── values_explorer/

    ├── __init__.py

    ├── data/

    │   └── loader.py

    ├── analysis/

    │   ├── __init__.py

    │   ├── clustering.py

    │   └── visualization.py

    └── utils/

        ├── __init__.py

        └── helpers.py

#+END_SRC

** Getting Started

1. Clone this repository

2. Install dependencies: ~pip install -e .~

3. Open exploration notebook: ~emacs notebooks/exploration.org~