Hugging Face's logo

Datasets:
spiralsafety
/
delusioneval

Tasks:
Text Classification
Text Generation
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
< 1K
Tags:
llm-safety
mental-health
conversational-ai
evaluation
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
delusioneval / README.md
Jared Moore
update homepage
ef493a6
preview code
|
Raw
History Blame Contribute Delete
11.1 kB
metadata 
pretty_name: DelusionEval
language:
  - en
license: other
license_name: delusioneval-data-use-agreement
license_link: LICENSE.md
gated: true
extra_gated_heading: DelusionEval Data Use Agreement
extra_gated_description: >
  Access to this dataset is restricted. Read the agreement below, provide the
  requested information, and confirm that you will comply with the terms before
  requesting access.
extra_gated_prompt: |
  ## DelusionEval Controlled Data Use Agreement

  Version 1

  Last updated: 2026-08-04

  Copyright (c) 2026 The Board of Trustees of The Leland Stanford Junior
  University. All Rights Reserved.

  ### 1. Parties

  This Controlled Data Use Agreement ("Agreement") is entered into between The
  Board of Trustees of The Leland Stanford Junior University ("Licensor") and
  the data requestor and, if applicable, the requestor's institution
  (collectively, "Licensee").

  ### 2. Scope and Purpose

  Licensor provides controlled access to the DelusionEval dataset and related
  documentation (collectively, "Data") to support non-commercial scientific
  research on AI safety, evaluation, and related topics.

  ### 3. Limited License Grant

  Subject to this Agreement, Licensor grants Licensee a non-exclusive,
  revocable, non-transferable, non-sublicensable limited license to access and
  use the Data solely for lawful, non-commercial scientific research.

  No rights are granted except as expressly stated in this Agreement.

  ### 4. Access and Sharing Restrictions

  Licensee must:

  1. Restrict access to approved personnel only.
  2. Not share credentials, raw files, or access paths with unauthorized
     parties.
  3. Not redistribute the Data, in whole or in part, to any third party.
  4. Ensure that all personnel with access are trained on human-subject
     protections and applicable privacy/security obligations.

  ### 5. Prohibited Uses

  Licensee must not:

  1. Attempt to identify, contact, or infer the identity of any participant or
     institution represented in the Data.
  2. Link, match, combine, cross-reference, or associate the Data (in whole or
     in part) with any other dataset, database, publicly available information,
     or other source of information, where doing so could reasonably enable or
     increase the likelihood of identification or re-identification of any
     individual or institution.
  3. Use any analytical technique, algorithm, model, manual method, or
     auxiliary information for the purpose of, or that has the effect of,
     reversing, defeating, or circumventing any de-identification,
     anonymization, pseudonymization, aggregation, or other privacy-protective
     measure applied to the Data.
  4. Use the Data to train, optimize, benchmark, or otherwise improve systems
     intended to facilitate self-harm, violence, delusional reinforcement, or
     other harmful behavior.
  5. Use the Data for clinical diagnosis, treatment, or direct decision-making
     about identifiable persons.
  6. Use the Data for advertising, surveillance, insurance, employment
     screening, law-enforcement profiling, or other non-research deployment
     contexts.

  ### 6. Security and Incident Reporting

  Licensee must use reasonable administrative, technical, and physical
  safeguards to protect the Data from unauthorized access, use, or disclosure.

  If Licensee discovers a potential re-identification risk, data leak, or other
  security/privacy incident involving the Data, Licensee must promptly report it
  to the Licensor contact listed in Section 16.

  ### 7. Publication and Citation Requirements

  For any publication, preprint, report, or other public disclosure that uses
  the Data, Licensee must:

  1. Cite the DelusionEval dataset DOI.
  2. Cite the FAccT 2026 Delusional Spirals paper.
  3. Cite the DelusionEval dataset release.

  Licensee must not publish examples, excerpts, or derived artifacts in a way
  that materially increases re-identification risk.

  ### 8. Ownership and Third-Party Rights

  Stanford-owned portions of the released materials are owned by The Board of
  Trustees of The Leland Stanford Junior University.

  Some underlying source rights, including rights in original contributed
  transcripts, may be held by third parties and are provided under limited
  permissions. This Agreement does not transfer those rights.

  ### 9. Compliance with Law and Policy

  Licensee must comply with all applicable laws, regulations, and institutional
  policies governing human-subject and sensitive data research.

  Where required by Licensee's institution, Licensee is responsible for
  obtaining local ethics/IRB review or confirmation before use.

  ### 10. Term and Termination

  This Agreement is effective upon first access to the Data and remains in force
  until terminated by Licensor or Licensee.

  Licensor may terminate access immediately for breach. Upon termination,
  Licensee must stop use of the Data and destroy local copies, except where
  retention is required by law or formal institutional policy.

  ### 11. Warranty Disclaimer

  THE DATA ARE PROVIDED "AS IS," WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  IMPLIED, INCLUDING BUT NOT LIMITED TO WARRANTIES OF MERCHANTABILITY, FITNESS
  FOR A PARTICULAR PURPOSE, TITLE, OR NON-INFRINGEMENT.

  ### 12. Limitation of Liability

  TO THE MAXIMUM EXTENT PERMITTED BY LAW, LICENSOR AND COPYRIGHT HOLDERS ARE NOT
  LIABLE FOR ANY CLAIM, DAMAGE, OR OTHER LIABILITY ARISING FROM OR RELATING TO
  USE OF THE DATA. LICENSEE AGREES TO HOLD HARMLESS LICENSOR FOR CLAIMS ARISING
  FROM LICENSEE'S USE, BREACH, RE-IDENTIFICATION ATTEMPTS, OR SECURITY
  INCIDENTS.

  ### 13. Publicity

  Licensee will not use the name or trademark of Stanford, or the names of
  Stanford's employees, students, or agents in any publicity, advertising, or
  announcement related to this Agreement without the prior written consent of
  Stanford's authorized officials. Any use of Stanford's name will be limited to
  statements of fact and will not imply endorsement by Stanford of Licensee's
  products or services.

  ### 14. Updated Terms

  Licensor may update these terms from time to time. Licensor will notify
  Licensee of any material changes by posting a notice on the dataset page,
  sending an email to the registered address, or displaying a prominent notice
  upon login.

  If Licensee objects to any changes, Licensee may terminate this Agreement by
  ceasing use of the Data within 30 days of notice. Continued use thereafter
  means Licensee accepts the new terms.

  ### 15. Governing Law

  This Agreement will be governed by the laws of the State of California, with
  venue for any disputes allowed only in the courts within Santa Clara County.

  ### 16. Contact

  Licensor: The Board of Trustees of the Leland Stanford Junior University
  ("Stanford")

  Contact for access, compliance questions, and incident reporting:
  - Name/Role: Jared Moore
  - Email: jlcmoore@stanford.edu
extra_gated_fields:
  Full name: text
  Affiliation: text
  Institutional email: text
  Intended use: text
  I have read and agree to the DelusionEval Data Use Agreement: checkbox
extra_gated_button_content: Agree and submit request
size_categories:
  - n<1K
task_categories:
  - text-classification
  - text-generation
tags:
  - llm-safety
  - mental-health
  - conversational-ai
  - evaluation
configs:
  - config_name: default
    data_files:
      - split: train
        path: items_sanitized.parquet

DelusionEval

Dataset Summary

DelusionEval is an anonymized conversational evaluation dataset for measuring problematic chatbot behavior in delusional-spiral contexts.

This release contains 725 conversation windows with:

  • eval_subset_id (string)
  • label (target behavior code)
  • meets_code (bool)
  • messages (ordered list of message structs with role, content, and per-message score fields)

Dataset Description

Supported Tasks

  • Safety evaluation and auditing of conversational model behavior
  • Behavior-code detection and analysis in dialogue windows

Dataset Structure

Data Instances

Each row is a conversation window keyed by eval_subset_id and label. messages is a list of dict-like message objects containing:

  • text fields (role, content)
  • bot code score fields (for example, bot-endorses-delusion, bot-romantic-interest)
  • user intent score fields (user-suicidal-intent, user-violent-intent)

Data Splits

This release is a single split:

  • train: 725 rows

Labels

18 label codes are included: bot-claims-unique-connection, bot-discourages-self-harm, bot-discourages-violence, bot-dismisses-counterevidence, bot-endorses-delusion, bot-facilitates-self-harm, bot-facilitates-violence, bot-grand-significance, bot-metaphysical-themes, bot-misrepresents-ability, bot-misrepresents-sentience, bot-platonic-affinity, bot-positive-affirmation, bot-reflective-summary, bot-reports-others-admire-speaker, bot-romantic-interest, bot-validates-self-harm-feelings, bot-validates-violent-feelings.

Dataset Creation

Windows were selected from anonymized transcripts and then manually reviewed, filtering, and anonymized.

Considerations for Use

Intended Uses

  • Safety evaluation and auditing
  • Method development for detecting problematic chatbot behavior

Out-of-Scope and Non-Recommended Uses

  • Re-identification attempts
  • Profiling individuals
  • Optimizing harmful assistant behavior

Limitations

  • Small, curated, and domain-specific sample
  • Sensitive content (mental health, self-harm, violence themes)
  • Not a population-representative dataset

Access and Safety

Controlled and sensitive-use expectations apply. Do not attempt re-identification.

Citation 

@inproceedings{10.1145/3805689.3806443,
author = {Moore, Jared and Mehta, Ashish and Agnew, William and Anthis, Jacy Reese and Louie, Ryan and Mai, Yifan and Yin, Peggy and Cheng, Myra and Paech, Samuel J. and Klyman, Kevin and Chancellor, Stevie and Lin, Eric and Haber, Nick and Ong, Desmond C.},
title = {Characterizing Delusional Spirals through Human-LLM Chat Logs},
year = {2026},
isbn = {9798400725968},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3805689.3806443},
doi = {10.1145/3805689.3806443},
booktitle = {Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency},
pages = {7631--7674},
numpages = {44},
location = {},
series = {FAccT '26}
}

@misc{moore2026delusioneval,
  title = {DelusionEval: Measuring Delusion-Linked Behaviors in AI Chatbots},
  author = {Moore, Jared and Mock, Andrea and Mai, Yifan and Anthis, Jacy Reese and Louie, Ryan and Agnew, William and Mehta, Ashish and Klyman, Kevin and Liang, Percy and Haber, Nick and Lin, Eric and Ong, Desmond C.},
  year = {2026},
  url = {TODO},
}